# Changelog Public, user-facing release notes for Deno Custom Nodes. This file intentionally stays short. Detailed engineering notes belong in private/local handoff notes, not the public changelog. ## Unreleased ## 0.7.105 - 2026-09-07 - MiniMax H3 Acc Loader now distinguishes a declared H3 conversion from other unsupported LoRA layouts and explains how to select an original Acc checkpoint. Guidance for compatible converted PDD LoRAs includes the required ComfyUI PDD support and built-in LoRA loader. ## 0.7.104 - 2026-09-06 - Added a DENO compatibility workaround for NVIDIA VFX output corruption at non-aligned widths, including the diagonal color bands reported with Megapixels sizing. RTX Video Super Resolution and its 2 Pass node now align the internal native output width and restore the exact requested output size. - Preserved Denoise/Deblur dimensions with temporary edge padding, including the first stage of 2 Pass. Existing node settings, saved workflows, batch order, output dtype/device, and already aligned processing paths are unchanged. ## 0.7.103 - 2026-09-05 - Moved URL image loading to urllib3's HTTP client, using its supported public-IP connection and original-host HTTPS verification options. Per-redirect address validation, direct connections, timeouts, response size limits, and saved workflows are preserved. - Supersedes v0.7.101 and v0.7.102, whose socket connection APIs were flagged by the Registry. Includes the fixes and internal optimizations listed under v0.7.101. ## 0.7.102 - 2026-09-05 - Replaced URL image loading's manual socket management with the standard-library connection helper. Validated public IPs, original HTTP Host and HTTPS certificate checks, redirect validation, timeouts, and saved workflows are unchanged. - Supersedes v0.7.101, which the Registry flagged for its direct socket creation pattern. Includes all v0.7.101 fixes below. ## 0.7.101 - 2026-09-05 - Preserved edits made while Ideogram Director view translation is pending, and restored surviving Visual Fold members when their representative node is deleted or an orphaned fold is reopened. - Fixed Local LLM Reviewer approval of negated phrases such as "isn't OK" and "not fully approved". Canceled Auto Retry requests now stop before submission, preserve unrelated manual runs, and restore an unsent seed change only when the user has not edited it. - Kept ComfyUI responsive during slow model-list, model-unload, and Director translation requests by moving provider I/O off the server event loop. - Matched Video Compare output to the preview: Toggle keeps the selected A/B side, and output duration follows the original input A's frame count and selected FPS, even after Swap. B is sampled over that duration; B becomes the timing source when A is absent. Save/encode nodes should use the same selected FPS. - Fixed Video Preview encoding with audio rates unsupported by AAC, including 192 kHz, by resampling to a supported rate while preserving timing. - Applied LTX Audio/Video LoRA strength correctly to diff and full-matrix LoKr patches without modifying cached adapters. Tiled samplers now pass through the original latent when the sigma schedule has no sampling steps. - Fixed URL image loading's DNS-rebinding gap by connecting to the validated public IP while preserving HTTPS hostname checks. URL loading now uses direct HTTP(S) connections and does not use environment proxies. - Reduced temporary memory use during image normalization and video preview encoding, and avoided redundant video-player text updates. Existing node IDs, input/output order, and saved workflow schemas are unchanged. ## 0.7.100 - 2026-08-31 - Extended `(Deno) Ideogram Director`'s existing `bboxes` output to connect to both `BBOX` and `BOUNDING_BOX` consumers, and exposed its active box count through a transient, non-serialized frontend bridge for `Ideogram4_MultiLora_BoundingBoxNode_Fedor`. Existing output order, saved `caption_data`, widgets, and BBOX links are unchanged. ## 0.7.99 - 2026-08-31 - Fixed `(Deno) Local LLM Loader` workflow loading on current ComfyUI frontend releases by deferring linked widget-socket cleanup until the node has an authoritative graph. Existing image, duration, audio-context, and prompt links are preserved through load, save, and reopen. - Removed duplicate global Local LLM wheel interception and synchronous layout reads. Local preview scrolling stays local, while normal canvas zoom and middle-button pan continue through ComfyUI's native interaction path. - Fixed `(Deno) Advanced Image Source Loader` input-folder thumbnails by trying compatible `/api/view` and `/view` URL variants before showing a fallback. Large input and external folders now render a bounded virtual grid with lazy thumbnails and preserved selection instead of creating every card at once. - Documented the existing opt-in route for calling an owned LM Studio server on a trusted LAN through the `Custom` provider and an exact `DENO_LOCAL_LLM_ALLOWED_HOSTS=IP:port` allowlist. The dedicated `LM Studio` provider remains localhost-only. ## 0.7.98 - 2026-08-29 - Followed up the Registry scanner compatibility cleanup by keeping the same four read-only environment lookups behind a direct standard-library import. This removes the scanner's remaining two false-positive source patterns without changing Local LLM host opt-in, Unsloth API-key lookup, SOS diagnostics, or any public node contract. ## 0.7.97 - 2026-08-29 - Added a one-time canvas migration for standard active UI workflows saved with the v0.7.92-v0.7.94 three-output `(Deno) MiniMax H3 Acc LoRA Loader`. Existing model links stay in place while former sampler and sigmas links move to editable stock `KSamplerSelect: euler` and `BasicScheduler: simple, steps: 8` nodes. Save the UI workflow once after it opens. Current single-output workflows, muted/bypassed nodes, unknown layouts, and malformed graphs are left untouched. Raw API prompt JSON does not run this frontend migration and must be exported again from a migrated UI workflow; files already saved after their legacy sampler/sigmas links were lost require manual reconnection. - Rephrased seven behavior-neutral source expressions that the Comfy Registry scanner mistook for environment manipulation, privilege escalation, socket binding, or a blocked URL. Node IDs, inputs, outputs, widgets, saved values, and runtime behavior are unchanged. ## 0.7.96 - 2026-08-27 - Improved `(Deno) MiniMax H3 Acc LoRA Loader` without changing its saved node contract: stock schedulers remain free to provide 4-12 or other descending sigma schedules, native ComfyUI `FinalLayer` modulation stays in control, and only the PDD video/audio projections are replaced. Full non-pruned INT8 models continue through ComfyUI's quantization-aware LoRA path. Curve-pruned models now automatically derive an AdaLN bridge from an installed matching full H3 checkpoint and rebase all 50 AdaLN updates; when no matching full checkpoint is available, the previous non-blocking compatibility fallback remains. ## 0.7.95 - 2026-08-27 - Changed `(Deno) MiniMax H3 Acc LoRA Loader` to return only the patched model and leave scheduler, sampler, and step controls to stock ComfyUI nodes. The loader now reads each sampling pass's actual descending sigma boundaries and dynamically fuses the checkpoint's 32 PDD intervals, including experimental 9/10-step and split-sigma latent-upscale paths. Simple/Euler at 8 steps remains the official recommended setup. Existing workflows must reconnect the former sampler and sigmas links. - Changed `(Deno) Text Encoder Unload` to use normal ComfyUI input caching instead of forcing every queue to invalidate its pass-through outputs. Unchanged preview sampling can now remain cached in Group Bypasser and similar preview-to-upscale workflows, while changed conditioning or CLIP paths still run the targeted unload. - Changed the MiniMax H3 native-version gate to a static compatibility import so the Registry scanner no longer mistakes the check for dynamic code execution. ## 0.7.94 - 2026-08-27 - Updated `(Deno) MiniMax H3 Acc LoRA Loader` to discover Alibaba Acc-LoRA files in both the normal ComfyUI `models/loras/` paths and the dedicated `models/minimax_h3_acc_loras/` paths, including configured extra-model roots. ## 0.7.93 - 2026-08-27 - Added curve-pruned MiniMax H3 compatibility to `(Deno) MiniMax H3 Acc LoRA Loader`. Comfy-Org `*_pruned_*` models now run by applying all compatible Alibaba LoRA updates and PDD heads while skipping only the 50 full-width AdaLN updates that cannot fit the pruned 8-wide curve basis; full models still receive the complete adapter. ## 0.7.92 - 2026-08-27 - Added `(Deno) MiniMax H3 Acc LoRA Loader` for direct use of Alibaba PAI's official FL2VA and Ref2VA Acc-LoRA/PDD safetensors with full native ComfyUI MiniMax H3 models. The node applies the complete adapter, automatically returns Euler and the trained 8-step sigma schedule, and rejects incompatible pruned models or schedules; LoRA weights and workflows are not bundled. ## 0.7.91 - 2026-08-24 - Reworked `(Deno) Text Encoder Unload` from the v0.7.90 wildcard `value` pass-through plus dependency-only `wait_for` into the typed prompt paths used by normal workflows: required `Positive Conditioning` and optional `Negative Conditioning`, with the latter accepting either an encoded negative prompt or `Conditioning Zero Out` and both paths available as named outputs. - Migrates saved v0.7.90 Text Encoder Unload conditioning links to the new named sockets when those workflows are reopened. The former wildcard node's non-conditioning uses are outside the new typed contract. The multilingual README, native help, and workflow screenshot now match the final node layout. ## 0.7.90 - 2026-08-21 - Added `(Deno) Text Encoder Unload`, an opt-in type-preserving workflow barrier that unloads only the explicitly connected CLIP/text encoder after all linked encoding branches finish, without globally unloading diffusion models, VAEs, or ControlNets. It includes current ComfyUI MatchType sockets, a legacy wildcard fallback, clear `--gpu-only` handling, and reliable per-queue execution. - Refreshed the public node catalog, multilingual setup guidance, native help, search terms, screenshots, Registry icon, and README banner, and retired the obsolete Bernini preview-backend update path now that current ComfyUI Stable includes native Bernini conditioning support. ## 0.7.89 - 2026-08-19 - Fixed `(Deno) Local LLM Loader` so Ollama JSON-schema output preserves clearly requested long-form word or character targets instead of collapsing them to a title or short introduction. The compatibility path keeps the original prompt, images, seed, cancellation, and unload behavior, while ordinary requests, source-length references, reviewer/rewrite workflows, LM Studio, and saved workflows remain unchanged. Addresses #77. ## 0.7.88 - 2026-08-15 - Added Unsloth Studio as a first-class `(Deno) Local LLM Loader` provider with authenticated model discovery and chat, Thinking control, manual and post-run unload, provider-specific URL restoration, and saved-workflow compatibility without storing the API key in workflows or PNG metadata. - Fixed LM Studio compatibility for models that do not expose reasoning configuration by retrying exactly once without `reasoning_effort` only before generation output starts, while preserving the prompt, images, seed, structured-output request, cancellation, and unload behavior. Fixes #68. ## 0.7.87 - 2026-08-13 - Fixed `(Deno) Resize Box` in `Keep Input Ratio` mode so connected VHS video previews provide their loaded frame ratio before execution, replacing the incorrect 1024 x 1024 fallback preview with the computed portrait or landscape output dimensions without changing backend sizing or saved workflow compatibility. ## 0.7.86 - 2026-08-12 - Added the complete official LTX 2.5 Distilled INT8 model set to `(Deno) Easy Model Download Helper`, including the two-stage x2 spatial upscaler, while preserving saved custom and legacy LTX 2.3 presets and keeping all gated downloads in the user's browser. - Updated `(Deno) LTX Sequencer` to request its image input lazily when guides are active and to follow current ComfyUI LTX guide metadata without changing its saved input or output schema. - Fixed the DENO input-folder browsers to skip external symlinks and junctions that ComfyUI cannot safely serve, with a clear route to `(Deno) Advanced Image Source Loader` for external folders. Fixes #71. - Fixed image-loader compatibility with Nodes 2.0 by keeping serialized path widgets hidden, making the Advanced loader gallery follow the available node height, and keeping disabled sources readable without rebuilding the whole gallery. Fixes #72. ## 0.7.85 - 2026-08-11 - Fixed `(Deno) Local LLM Loader` so, whenever the node executes, its final Result is embedded in that output's workflow metadata and restored inside the node when a saved PNG or workflow is reopened, without persisting Thinking/reasoning content. ## 0.7.84 - 2026-08-11 - Added `(Deno) Audio Transcript` for optional local Whisper transcription with user-supplied wording priority, segment timing, confidence context, official checkpoint caching, audio passthrough, and post-run model unload. - Added `(Deno) Audio Analysis Finalizer` to keep only the documented acoustic-analysis fields from ComfyUI `TextGenerate` output and optionally unload the analysis CLIP model. - Added an optional `audio_context` STRING input to `(Deno) Local LLM Loader` so transcript and acoustic evidence can guide prompt direction as data without replacing the user's prompt. - Added a portable beginner MiniMax H3 R2V audio-reference workflow that uses the stock H3 reference-audio path and muxes the selected source-audio section unchanged into the final video. ## 0.7.83 - 2026-08-09 - Hardened `(Deno) Local LLM Loader` so Ollama and LM Studio return only the validated final answer while keeping reasoning and provider-added preambles out of the Result, without changing saved prompts, images, seed controls, or arbitrary multiline and Reviewer JSON outputs; LM Studio now applies the selected seed to generation requests. - Added one bounded finalization attempt for completed malformed or thinking-only replies, then fail closed instead of forwarding unvalidated model text; LM Studio generation now uses its official JSON Schema chat-completions path while retaining native model discovery and unload. ## 0.7.82 - 2026-08-09 - Fixed `(Deno) Local LLM Loader` with Ollama Thinking enabled so a completed reasoning-only response automatically continues once to produce the final answer, while preserving the original prompt, images, seed, cancellation, and unload behavior. ## 0.7.81 - 2026-08-08 - Fixed `(Deno) Local LLM Loader` so a vertically enlarged prompt area can shrink again, follows ComfyUI's native DOM layout, and keeps its compact size after saving and reopening a workflow. ## 0.7.80 - 2026-08-07 - Fixed `(Deno) Local LLM Loader` so LM Studio reasoning-capable models receive an explicit `reasoning: off` request whenever Thinking is disabled, including after the model-list cache expires. - Removed the automatic second full generation after an empty LM Studio stream; the node now fails immediately and records the requested and actually applied reasoning modes accurately. ## 0.7.79 - 2026-08-06 - Changed `(Deno) MiniMax H3 Multi Reference Image Loader` preview cards to follow each source image's aspect ratio, keeping mixed landscape and portrait references fully visible without preview cropping. ## 0.7.78 - 2026-08-06 - Improved `(Deno) Local LLM Loader` with applied System Prompt preset status, an optional `video seconds` FLOAT socket, and workflow-tab model restoration that waits for a successful model-list refresh before declaring a saved model missing. - Fixed DENO Floating Tools so the first-aid error icon returns to normal after the matching retry succeeds without being reactivated by an old ComfyUI error toast. ## 0.7.77 - 2026-08-06 - Added durable per-user system-prompt presets to `(Deno) Local LLM Loader`, including a one-time browser-preset import when legacy presets are detected. Fixes #54. - Kept frozen audio exact in both outputs of `(Deno) LTX High resolution Tiled Sampler` when external samplers such as RES4LYF `res_2s` introduce finite internal audio drift, while retaining structural and non-finite-value validation. Fixes #55. ## 0.7.76 - 2026-08-06 - Added aspect-ratio-locked zoom cropping to `(Deno) Resize Box`: drag the crop area to reposition it or drag any corner to zoom while keeping the selected output megapixels and dimensions fixed. Crop position and zoom persist with the workflow. - Updated `(Deno) LTX Tiled Spatial Upscaler` for ComfyUI 0.30's managed `ModelPatcher`/DynamicVRAM upscaler contract while preserving the legacy raw-module path. Fixes #56. - Updated `(Deno) LTX High resolution Tiled Sampler` to accept both packed and nested AV callback latents while retaining strict denoised-video shape and frozen-audio checks. Fixes #59. - Clamped `(Deno) Video Compare` slider writes and repaired already-saved out-of-range split positions. Thanks to @NubeBuster for PR #57. ## 0.7.75 - 2026-08-05 - Added draggable crop positioning to `(Deno) Resize Box`, including an Ideogram Director-style connected-image loader and a semi-transparent preview clipped to the exact output frame, while preserving the existing center-crop default, output-size corner drag, and saved-workflow behavior. - Added a mixed-size `image_list` output to the MiniMax H3 reference-image loader so the same ordered references can connect directly to `(Deno) Local LLM Loader` without duplicating image-loader nodes. ## 0.7.74 - 2026-08-05 - Added a one-cable MiniMax H3 multi-reference image loader that preserves each source image's decoded dimensions and order, plus a stock-compatible H3 Reference to Video wrapper that keeps native video and audio Autogrow inputs. - Added a complete portable MiniMax H3 reference-to-video example workflow and Registry discovery metadata for the new nodes. - Restored legacy saved-workflow image paths in the standard Multi Image Loader while keeping the H3 loader restricted to ComfyUI input files. - Kept canvas wheel zoom and middle-button pan available over both image-loader panels, and synchronized their inner galleries with manual node resizing and save/reopen state. ## 0.7.73 - 2026-08-03 - Fixed `(Deno) Resize Box` Keep Input Ratio live preview after ComfyUI changed newly created canvas node IDs to strings, while preserving numeric-ID workflows and backend sizing behavior. ## 0.7.72 - 2026-07-25 - Added per-element enable/disable controls and optional connected Summary/Background overrides to `(Deno) Ideogram Director` while preserving saved element order, values, and disconnected board behavior. - Opened BBOX editing beside the clicked region, kept the editor inside the visible node or fullscreen area as its contents resize, and added repeated `Alt`+click selection for overlapping BBOX regions without moving or reordering them. - Added `llama-swap` as a first-class `(Deno) Local LLM Loader` provider with OpenAI-compatible generation, live running-state checks, manual and post-run unload controls, saved-workflow compatibility, and server-owned timeout guidance. - Removed the synchronous LM Studio model-list request that previously ran before generation, so repeated warm-model runs can start without waiting for metadata. - Removed machine-specific preview, workspace, and model-root state from bundled public workflows so the examples stay portable across ComfyUI installations. ## 0.7.71 - 2026-07-22 - Fixed `(Deno) LTX Sequencer` growing to its full 50-slot height after a workflow reload, workflow-tab switch, or browser-tab restore when the node had been manually resized, and made already-affected saved workflows recover their compact size automatically the next time they load. ## 0.7.70 - 2026-07-19 - Changed DENO Floating Tools so its `NEW` badge and version row track ComfyUI Stable only, without checking or displaying separate frontend and workflow-template package versions. ## 0.7.69 - 2026-07-18 - Preserved saved workflows that used the retired video-only LTX tiled sampler by mapping them through ComfyUI's native node replacement flow while keeping current LTX AV sampler defaults, validation, outputs, and visible controls unchanged. - Fixed Local LLM cleanup and control races so failed or stopped runs still release an owned model once, late Refresh responses cannot overwrite a newer run, Stop and Unload requests finish safely, and negative reviewer verdicts are not mistaken for approval. - Fixed workflow-scoped Local LLM Reviewer snapshots, and hardened Video Preview file replacement and cleanup so parallel or saved previews do not overwrite or remove each other's state. - Fixed Video Compare batch alignment, fully disabled RTX two-pass passthrough, and exact LTX Sequencer no-op handling without changing results for already aligned or enabled workflows. - Hardened saved LoRA values, Bernini choices, Ideogram save paths, update/help caches, preview file replacement, and canvas cleanup so invalid or stale state fails clearly without changing normal controls or outputs. ## 0.7.68 - 2026-07-18 - Fixed `(Deno) Ideogram Director` result-image restoration, panel cleanup, reroute backdrop lookup, queue preflight handling, and targeted pending-import events without changing normal generation settings or outputs. - Fixed `(Deno) Local LLM Loader` and `(Deno) Local LLM Reviewer` workflow isolation, progress routing, queue-time seed updates, popup ownership, prompt preservation, and scheme-less local server URL restoration. - Fixed multi-image, advanced image-source, and video-preview handling so active source failures stop clearly, unchanged local files avoid repeated full validation, folder changes invalidate correctly, and temporary previews from different workflows no longer collide. - Fixed Multi LoRA and LTX Multi LoRA legacy saved boolean values so disabled slots remain disabled without changing current widget order or normal loading behavior. - Fixed Resize Box disconnected size reporting and Multi Image Loader-to-Sequencer reroute notifications while preserving the existing preview drag feel and Sequencer polling behavior. - Corrected Local Model Downloader documentation to match its existing registered-root lookup behavior. ## 0.7.67 - 2026-07-03 - Fixed `(Deno) Local LLM Loader` and `(Deno) Local LLM Reviewer` saved-workflow handling so provider URLs, prompts, thinking state, and reviewer state stay in the correct saved slots after workflow save, reload, copy, and provider changes. - Kept Local LLM server access local-first by default while allowing advanced users to opt in to a specific trusted LAN server with `DENO_LOCAL_LLM_ALLOWED_HOSTS=LAN-IP:port`. - Improved `(Deno) LTX High resolution Tiled Sampler` guide and mask conditioning stability with stricter spatial conditioning checks, clearer guide token-count mismatch errors, and pixel-mask crop validation. ## 0.7.66 - 2026-07-01 - Fixed DENO Floating Tools Error Help so large ComfyUI failures keep the browser responsive, the report is prepared only when the user opens it, and the error icon returns to normal after the next confirmed idle run. ## 0.7.65 - 2026-06-30 - Added `Error Help` to DENO Floating Tools. When a ComfyUI run fails, the helper highlights the DENO icon and opens a GPT/Gemini-ready report window with workflow, Python environment, package, GPU, custom-node, log, and traceback details. Sensitive tokens, cookies, passwords, private keys, and URL credentials are redacted before copy. - Improved `(Deno) Ideogram Director` with optional width, height, and megapixel inputs, plus a Generate target selector for running one intended output branch while keeping the default all-output behavior. - Improved `(Deno) Local LLM Loader` state handling for copy/paste and saved workflows, and raised image attachment detail for local vision models while keeping uploads bounded. - Improved `(Deno) LTX High resolution Tiled Sampler` guide-aware handling and in-app docs for LTX AV guide workflows. ## 0.7.64 - 2026-06-27 - Restored release CI compatibility for environments that expose a lightweight torch stub without tensor operations, keeping the 0.7.63 Floating Tools and final LTX AV tiled node changes intact. ## 0.7.63 - 2026-06-27 - Improved DENO Floating Tools update checks so the panel refreshes live installed versions from the running ComfyUI instead of showing stale cached values after external updates. - Replaced the old video-only LTX tiled sampler registration with `(Deno) LTX AV Step-Fused Tiled Sampler`, which uses the full audio latent as video context while keeping audio frozen. - Kept `(Deno) LTX Tiled Spatial Upscaler` as the tiled high-resolution upscaler and changed the final LTX tiled nodes to `2 x 2` fresh defaults with extra memory cleanup enabled. ## 0.7.62 - 2026-06-24 - Fixed `(Deno) Local LLM Loader` so the Thinking and Result preview panels stay inside the node instead of stretching across nearby nodes. ## 0.7.61 - 2026-06-24 - Fixed `(Deno) Multi Image Loader` so running with no selected images now stops with a clear error instead of producing a placeholder output. - Stabilized `(Deno) LTX Sequencer` dynamic timing rows so compact layouts, saved workflows, linked high-index rows, copy/paste, and manual node resizing restore more reliably. ## 0.7.60 - 2026-06-23 - Fixed Easy Model Download Helper model-root detection for ComfyUI setups with more than one registered model folder. - Selecting a model root now changes the displayed/copy target path without hiding files that ComfyUI can already find through another registered model path. - Kept model discovery limited to ComfyUI-registered model paths instead of probing nearby folders. ## 0.7.59 - 2026-06-23 - Added `[BETA] (Deno) LTX Tiled Spatial Upscaler` for tiled video-latent spatial upscale experiments in high-resolution LTX workflows. - Added `[BETA] (Deno) LTX Step-Fused Tiled Sampler` for video-only tiled second-pass sampler experiments. - Improved the Visual Fold toolbar buttons so Fold and Align appear as readable DENO badges with hover descriptions while staying inside ComfyUI's native selection toolbar. ## 0.7.58 - 2026-06-23 - Smoothed Visual Fold and Align integration with ComfyUI's multi-selection toolbar so the DENO buttons appear with the native toolbar and stay stable while native menus are open. ## 0.7.57 - 2026-06-22 - Fixed Ideogram Director Regenerate so a connected JSON prompt no longer overwrites manual box and description edits when the same prompt was already applied. ## 0.7.56 - 2026-06-22 - Fixed Visual Fold on Desktop so folded nodes with DOM text widgets no longer leave prompt boxes visible under the folded chip. ## 0.7.55 - 2026-06-22 - Fixed Visual Fold and Align so their toolbar integration no longer blocks ComfyUI's own multi-selection floating controls. ## 0.7.54 - 2026-06-21 - Fixed Ideogram Director bbox editing so drawing, moving, and resizing boxes keeps responding inside the node on more ComfyUI browser/runtime setups. - Fixed Ideogram Director's board preview so changing output resolution updates the canvas shape immediately even when a generated image is already shown. ## 0.7.53 - 2026-06-21 - Added a Local LLM Loader Tip popup that shows how to chain LLM nodes for prompt generation, review, branching, and final cleanup. - Fixed the Tip button drawing so it does not affect ComfyUI's normal widget text alignment. ## 0.7.52 - 2026-06-20 - Polished DENO Floating Tools update display: the update badge now stays inside the icon area, the icon stays still, and the panel uses shorter English update copy. ## 0.7.51 - 2026-06-20 - Improved Ideogram Director's Elements panel: crowded region lists can be widened, rows show the actual element text, and double-click editing is easier when boxes overlap. - Made Copy JSON more reliable by adding fallback copy paths and a manual copy dialog when the browser blocks clipboard access. - Cleaned the public GitHub surface so internal DENO operating notes stay local-only while user-facing docs remain public. ## 0.7.50 - 2026-06-19 - Added optional DENO Floating Tools, a small draggable helper that can free ComfyUI VRAM and show read-only Portable update status when enabled. - Documented the new three-surface verification baseline: Portable ComfyUI first, then official Desktop, then Easy-Install/EZi Desktop mode. ## 0.7.49 - 2026-06-19 - Fixed ComfyUI EZi/Desktop startup hangs caused by Easy Model Download Helper checking model folders while ComfyUI was still loading its node list. ## 0.7.48 - 2026-06-19 - Improved the DENO node info button: update details now open only on click, the popup closes when users click or wheel outside it, and the update card shows short release-note bullets without extra rollback clutter. - Fixed Ideogram Director's fullscreen Language picker so Escape closes it consistently, matching the on-screen `Esc to close` hint. ## 0.7.47 - 2026-06-19 - Fixed Local LLM Loader validation for queued/list-wrapped saved combo values, so current and legacy seed, memory, and ComfyUI model-unload settings validate correctly before execution. - Improved Visual Fold drag handling so Fold/Rename/Align controls stay hidden while nodes or groups are being dragged, and focus changes such as Alt+Tab cannot leave the controls stuck hidden. ## 0.7.46 - 2026-06-18 - Hardened saved-workflow compatibility for stale combo values across prompt guides, Local LLM nodes, Ideogram Director, resize/image utilities, video compare, and RTX VFX helpers. - Disabled, hidden, off, or inactive saved options from missing drives or older workflows no longer block ComfyUI before the node can ignore them. Active missing options still stop with a clearer field-specific message. ## 0.7.45 - 2026-06-18 - Fixed Multi LoRA and LTX Multi LoRA saved-workflow validation so disabled LoRA slots no longer block execution when their saved file is on a removed external drive or USB. Enabled missing LoRAs still stop with a clear slot-specific message. - Added Resize Box regression coverage for Keep Input Ratio so landscape source tensors keep the expected orientation. ## 0.7.44 - 2026-06-18 - Improved Ideogram Director bbox editing ergonomics: tiny boxes are easier to drag from the number badge, stale pixel-size popups no longer cover the board, and board undo/redo stays on the visible `↶` / `↷` buttons without taking over ComfyUI's graph undo. ## 0.7.43 - 2026-06-18 - Preserved saved LoRA selections in Multi LoRA and LTX Multi LoRA workflows even when a saved LoRA file is missing from the current PC's dropdown list. - Preserved saved RTX VFX Easy Upscale device selection instead of resetting it during frontend setup. - Preserved Local LLM Reviewer approve-once review state across Save, F5, and reopen. - Hid Visual Fold floating controls while nodes or groups are actively being dragged, so the Fold button no longer appears mid-drag. - Hardened Ideogram Director resolution handling so arbitrary image-analysis sizes such as `1712:880` do not silently replace the user's current output size, and restored saved custom sizes show a consistent megapixel value when the size popup opens. ## 0.7.42 - 2026-06-17 - Fixed Visual Fold selection handling so stale ComfyUI selection flags no longer leave Fold controls floating after a blank-canvas click or a one-node selection. - Improved Local LLM Loader saved-workflow restore so the visible `Thinking` toggle survives Save, F5, and reopen for current Ollama layouts. - Updated Local LLM Loader `Thinking` and `Result` `More` popups so they keep following live streaming text while open without forcing the scroll position when the user is reading older text. - Polished Ideogram Director's language refresh button so saved workflows reopen with the button at its normal width instead of a narrow vertical mark. - Hardened saved-workflow migration for Bernini Prompt Guide and RTX VFX Video Finisher so legacy generated-widget layouts preserve the user's visible values. ## 0.7.41 - 2026-06-17 - Added pack-wide ComfyUI Info panel descriptions for all public Deno nodes, including input and output explanations. - Improved the Deno node info button so it shows the installed pack version, checks Comfy Registry, and marks available updates with a yellow `i` plus a small `!` badge. ## 0.7.40 - 2026-06-17 - Improved Ideogram Director Elements ordering: the right-side list now reads visually front-to-back, shows a green insertion line while reordering, and keeps each box's editor color attached to that box after reordering. - Added Ideogram Director board undo/redo buttons beside Copy/Paste/Clear for users who prefer visible controls over Ctrl+Z/Ctrl+Y. - Added an Ideogram Director language refresh button so the current board can be translated again after loading a layout or pasted JSON. It reuses the saved translation engine and fallback dialog. - Protected legacy TEXT boxes whose literal rendered word is stored only in `desc` during board-view translation. ## 0.7.39 - 2026-06-17 - Improved Ideogram Director language translation recovery: if Google is blocked or unreachable, the node now explains the network/region issue, lets users switch to MyMemory, LibreTranslate, or a custom LibreTranslate server, and stops Generate/Copy JSON instead of silently passing a non-English prompt downstream. - Kept Ideogram Director TEXT box words protected during translated editing and final English output conversion. - Shortened the Ideogram Director top-bar layout button to `Layouts` so it stays readable at normal node widths. ## 0.7.38 - 2026-06-16 - Added a Deno Custom Nodes banner for the ComfyUI Manager and Registry listing. - Added Ideogram Director `Language` view so board descriptions can be read and edited in another language while final output stays model-ready English. - Improved tiny or overlapping Ideogram Director bbox editing by making the number badge a reliable drag handle. - Fixed Ideogram Director's resolution popup so it opens as a viewport overlay instead of being clipped inside the node. ## 0.7.37 - 2026-06-16 - Fixed Local LLM Loader `Seed Mode` so `increment`, `decrement`, and `randomize` update the visible seed after each queued run without shifting saved workflow widget values. ## 0.7.36 - 2026-06-16 - Fixed Ideogram Director's Desktop/Recreate-node sizing path so the editor recovers to a usable board instead of collapsing into a narrow rail or clipping Generate. - Fixed LTX Model Loader GGUF rows so saved and fresh nodes keep the correct model/VAE/text-encoder mapping and preserve external model paths across refresh. - Fixed LTX Prompt Guide so saved positive and negative prompt text survives Save, F5, and workflow reopen. - Fixed Local LLM Loader so missing saved Ollama/LM Studio models are shown as unavailable on other PCs, and LM Studio reasoning payloads now respect each model's supported options. ## 0.7.35 - 2026-06-16 - Fixed Ideogram Director so a manually enlarged node can be shrunk again with the ComfyUI resize handle. - Restored mouse-wheel scrolling inside Ideogram Director's right prompt/style/elements panel while keeping the board canvas-first. ## 0.7.34 - 2026-06-16 - Fixed Ideogram Director so ComfyUI resize/fit interactions no longer collapse the board to about half height. - Preserved user-resized Ideogram Director node sizes while keeping saved workflows compatible. ## 0.7.33 - 2026-06-15 - Added Ideogram Director, a visual Ideogram 4 JSON/bbox prompt builder with editable boxes, layout presets, style presets, and model-ready prompt output. - Improved Ideogram Director's connected JSON prompt flow so existing boards ask before replacement, invalid JSON is rejected clearly, and current-board edits can continue safely. - Kept the standalone Translator and Random Prompt Box out of the public release while preserving Ideogram Director's built-in Translate On/Off helper. - Hardened public workflow migration checks for older DENO workflow files. ## 0.7.32 - 2026-06-14 - Added Prompt Only final prompt extraction for Local LLM Loader, so Ollama and LM Studio models that add analysis text can pass only the final prompt downstream. - Models that cannot follow the Prompt Only format now stop with a clear error instead of sending analysis text into the workflow. ## 0.7.31 - 2026-06-14 - Hardened Local LLM Loader's local Ollama / LM Studio HTTP path so non-local URLs are rejected before any connection is opened. ## 0.7.30 - 2026-06-13 - Fixed release test compatibility for the Local LLM Loader and Reviewer package validation. ## 0.7.29 - 2026-06-13 - Added Local LLM Loader for local Ollama and LM Studio prompt workflows, with separate Stop LLM and Unload LLM controls. - Added Local LLM Reviewer for gating IMAGE and AUDIO save paths from review text, including one-shot approval and reviewer-side rerun controls. ## 0.7.28 - 2026-06-03 - Improved ComfyUI Manager, Comfy Registry, and GitHub discovery metadata for Bernini Prompt Guide, Bernini conditioning helpers, Wan2.2, reference video edit, and prompt guide searches. - Updated the Bernini preview backend update BAT to avoid Windows delayed-expansion parsing failures during the real update path. ## 0.7.27 - 2026-06-03 - Added Bernini Prompt Guide for KJ-style Bernini prompt prefixes, readable System Prompt mode labels, automatic reference-image prompt hints, a collapsible negative prompt section, and Wan2.2 negative preset autofill. - Added a Bernini preview backend update BAT for test portable ComfyUI folders while the upstream Bernini backend is still a draft PR. ## 0.7.26 - 2026-06-01 - LTX Model Loader keeps saved model selections during ComfyUI refresh instead of falling back to `__none__`. - Multi Image Loader now stops with a clear error when selected images are missing or unreadable, and refreshes correctly when selected image files change. ## 0.7.25 - 2026-05-31 - Easy Model Download Helper no longer appears twice in the node list; older workflow IDs are handled as a migration instead. - Added the public LTX 2.3 8GB VRAM workflow as a compatibility baseline for future DENO node updates. ## 0.7.24 - 2026-05-31 - Older workflows that still contain the previous LTX 8GB download helper node now open normally instead of showing an UNKNOWN missing-node box. ## 0.7.23 - 2026-05-30 - LTX Model Loader now restores older workflow dropdown values correctly and prevents hidden inactive model fields from blocking Checkpoint, KJ, or GGUF runs. ## 0.7.22 - 2026-05-27 - Video Preview shows a compact current-video info badge with resolution, FPS, frame count, and duration. - LTX Model Loader model dropdowns now hide recommended files that are not actually installed and avoid auto-selecting unrelated models. ## 0.7.21 - 2026-05-27 - Video Preview, Video Compare, and Image Compare preserve user-resized node sizes instead of repeatedly snapping back to media auto-fit. - Video Preview keeps hover-to-hear audio active when a new preview finishes loading under the cursor. ## 0.7.20 - 2026-05-26 - RTX VFX node panels keep ComfyUI canvas wheel and middle-click navigation available. - Visual Fold no longer shows Fold Group from stale group selection while normal nodes are selected. ## 0.7.19 - 2026-05-26 - Video Compare output naming polish. - RTX VFX upscale sizing now supports exact video sizes without forced 32px rounding. - Public changelog and release-note workflow added.
Previous Public Highlights ### 0.7.18 - 2026-05-24 - Multi Image Loader path-copy reliability update. - LTX checkpoint-style loader UI compatibility update. ### 0.7.17 - 2026-05-22 - Added DENO Visual Fold workflow cleanup tools. - Added the generic Multi LoRA Loader. - Added LoRA row ordering controls. ### 0.7.10 - 2026-05 - Added RTX VFX helper and installer flow polish.
## Release Note Style - Keep each public entry short and outcome-focused. - Prefer what users can see or benefit from. - Avoid file-level or implementation-level detail here. - Put technical investigation, verification notes, and local runtime details in private/local handoff notes.