v2.3.1 - Ideogram 4 structured-control recipe corrected

Fixed

• Ideogram 4 structured-control nodes restored (they are real). v2.3.0 wrongly claimed the
  Create Bounding Boxes -> Build JSON Prompt (Ideogram) flow had "no dedicated bounding-box or JSON-builder
  node." Those nodes exist in current ComfyUI core: CreateBoundingBoxes (comfy_extras/nodes_bounding_boxes.py;
  outputs preview / bboxes (BOUNDING_BOX) / elements (ARRAY)), BuildJsonPromptIdeogram ("Build JSON
  Prompt (Ideogram)", nodes_json_prompt.py; the ARRAY plus high_level_description / background / style /
  aesthetics / lighting / medium / color_palette COLORS -> a DICT caption), and Ideogram4Scheduler
  (nodes_ideogram4.py). The v2.3.0 note was written after checking only nodes_ideogram.py (V3 / V4 only)
  against a local core (0.25.1) that predated these modules, and no template uses them yet; the
  CreateBoundingBoxes widget landed in frontend v1.48.2 (2026-07-11). MODELS.md now documents the real
  buildable chain (CreateBoundingBoxes -> BuildJsonPromptIdeogram -> Ideogram 4), confirmed from the node
  source. The Grok 1K / 2K, ModelMergeKrea2 and quant_int8_convrot.py corrections from v2.3.0 stand.

v2.3.0 - Seedream 5.0 Pro and Seed Audio recipes, OCIO v1.2.2, native INT8-ConvRot

Added

• Seedream 5.0 Pro (ByteDance) - new official recipe. ByteDance's latest image model: multi-modal in ONE node
  (ByteDanceSeedreamNodeV2 - text-to-image + precise image editing + multi-image inputs), strong character /
  product consistency, region-precise editing, and structured layouts (legible small text); official templates
  api_bytedance_seedream_5_0_pro_{t2i,image_edit}. Shares the Seedream 5.0 CoT prompt style (no quality boosters).
• Seed Audio 1.0 (ByteDance) - new official recipe. A new audio model: generates a full audio SCENE (sound
  design + named-character dialogue + background music) in one pass, three modes t2a / ta2a / ti2a
  (ByteDanceSeedAudio -> SaveAudioAdvanced); official templates api_bytedance_seed_audio1_0_{t2a,ta2a,ti2a}.
• INT8-ConvRot is native in ComfyUI v0.27.0 (ADVANCED.md refresh). ComfyUI v0.27.0 shipped native int8-convrot
  weight quantization - the ConvRot rotation-based method (arXiv 2512.03673): ~half the FP16 size, faster than
  FP16, matches or beats FP8, faster + cleaner than FP8 on Turing / Ampere, and small enough to bring big models
  to 8-12 GB (and Pascal) cards (LTXV 2.3 1920x1088: 268 s -> 140 s). Refreshed the ADVANCED.md INT8 note (native
  path + the ConvRot paper + the comfy-model-tools quant_int8_convrot.py converter + Comfy-Org's HF weight
  uploads for Wan 2.2 Animate / Z-Image / SeedVR2; extends to INT4, covers DiT / LLM / multimodal / UNet), and
  updated the Flux2-Klein entry (its INT8-ConvRot quant now loads natively). The third-party ComfyUI-INT8-Fast
  is now fully superseded.
• Comfy-MSS - music source separation pack (NODE_LIBRARY/audio.md). Documented pymss-project/comfy-mss
  (ComfyUI nodes for the pymss library): split an AUDIO stream into instrument / vocal stems via MSS and
  VR/UVR models, with ensemble / normalize / phase-invert utilities - a fuller-featured alternative to the
  single-purpose MelBandRoFormer stem step.
• ComfyUI-OCIO reference bumped to v1.2.2 (NODE_LIBRARY/ocio.md + ADVANCED.md + README). No node changes
  since v1.2.0 (still nine nodes on the native VIDEO wire); v1.2.1 / v1.2.2 add a reproducible Dockerized CI
  round-trip test and an honest accuracy write-up. Corrected the accuracy story: the headline is per-transform
  bit-exact parity (0.000e+00), the end-to-end ACEScg -> LogC -> Rec.709 -> back round-trip is 4.5e-6
  max (single-precision LUT residual - NOT bit-for-bit lossless, but ~100x below a half-float EXR delivery grid),
  and the histogram chart is a distribution-shape sanity check, not an accuracy proof. The Docker / CI round-trip
  harness is credited to Sam Hodge (PR #1); the nodes + accuracy suite are Slava Sexton's. Also fixed the
  stale "eight nodes" README / BUILDING_NODES count to nine (OCIO Player shipped in v1.2.0).
• ComfyUI v0.27.0 coverage swept. Beyond INT8-ConvRot: added an Ideogram 4 JSON-prompt note to the Ideogram
  entry (write the IdeogramV4 prompt itself as a structured JSON-style caption; the official api_ideogram_v4_t2i
  template builds it with a GeminiNode magic-prompt - there is no dedicated bounding-box or JSON-builder node),
  a Grok Image 1K / 2K resolution note, and the core ModelMergeKrea2 block-merge node. Verified the rest of
  v0.27.0 is already in the kit (Seedance 2.0 Mini + 4K, HappyHorse 1.1, the Nano Banana 2 Lite + Gemini Video Omni
  partner nodes, and the v0.27.0 INT8 / frontend-Manager bugs already in KNOWN_ISSUES).
• Instruction editing on Krea 2 (community, experimental). Added to the Krea 2 entry: ostris/ComfyUI-Krea2-Ostris-Edit
  (Ostris / AI Toolkit - a Text Encode + Model Patch node pair that feeds reference latents into Krea 2, a
  text-to-image model, so it can edit an input image) and the community edit LoRA
  reverentelusarca/krea2-detail-enhancer-edit-lora (a detail enhancer, trigger "enhance this image",
  krea2-community-license). Flagged with the author's own honest caveats - not Flux.2 Klein / Qwen-Image-Edit
  precision (alters the image, shifts lighting/color, can fault on horizontal aspect ratios).

Count change: 69 -> 71 recipes, 150 -> 152 models, 545 -> 549 templates - from the two new official ByteDance
recipes (Seedream 5.0 Pro image + Seed Audio 1.0 audio) picked up by the Saturday weekly-update check; INT8-ConvRot,
Comfy-MSS, the v0.27.0 doc sweep and the Krea 2 edit method do not change the counts. All count places synced
(README, MODEL_INDEX, and the cover / models_by_modality / architecture / templates banners re-rendered; chip now
152 models · 71 recipes).

v2.2.0 - OCIO v1.2.0 video nodes, node-authoring lessons, new recipes

Added

• General ComfyUI node-authoring lessons from the ComfyUI-OCIO v1.2.0 video pipeline. Folded 9 universal,
  pack-agnostic lessons into BUILDING_NODES.md (+ VIDEO specifics in NODE_LIBRARY/video.md): the
  mutually-exclusive IMAGE-or-VIDEO dual socket (VideoFromComponents), the reload-safety app.configuringGraph
  guard (a mishandled onConnectionsChange wipes the whole graph's links), why an OUTPUT slot cannot be safely
  hidden, the Vue-nodes RETURN_NAMES fix, the servable-temp-file video-preview contract, VALIDATE_INPUTS for
  renamed combo values, the Windows temp-file ffmpeg decode (12-25x faster than a subprocess pipe), the on-node
  float/HDR WebGL viewport pattern, and the limits of scripting a live-ComfyUI screenshot.
• ComfyUI-OCIO reference refreshed to v1.2.0 (NODE_LIBRARY/ocio.md + the ADVANCED.md mention). Was pinned at
  v1.0.1 (eight nodes); now documents nine nodes, the native ComfyUI VIDEO pipeline (mutually-exclusive
  IMAGE-or-VIDEO sockets on all nine), the new OCIO Player on-node viewport, and the bit-exact accuracy proof
  (0.000e+00 max-abs error across 9 transforms x 4 fixtures). Historical CHANGELOG entries left untouched.
• LumiPic - single-image SDR -> HDR LoRA. Documented oumoumad/LumiPic (MIT) under Qwen-Image-Edit in
  MODELS.md: the IMAGE analog of the LTX-2.3 HDR IC-LoRA (same LumiVid paper, arXiv 2604.11788), a LoRA that emits
  an ARRI-Log-encoded frame decoding to scene-linear HDR EXR. Covered the three bases (Qwen-Image-Edit-2511
  mature; Flux.2 Klein 4B / 9B alpha), the LogC3 (ceiling ~55) vs LogC4 (V10 alpha, ceiling ~470) curves and which
  checkpoints to pick, the ComfyUI route (ready HF graphs + the ComfyUI_Gear decode node), and the honest V10
  caveat (Qwen V10 gain shows in diffusers but not yet in ComfyUI; klein4b_v10_logc4_step1500 is the one that
  holds up). Cross-linked from the LTX-2.3 HDR entry and ADVANCED.md.
• ComfyUI_Gear (oumad) node pack. Added to NODE_LIBRARY/custom-author.md: LogC3 / LogC4 Decode + Save EXR
  (also decodes our LTX-2 HDR IC-LoRA) and a full ACEScct Color Grade panel. Flagged the honest relation to our
  ComfyUI-OCIO - Gear decodes but keeps the source primaries, so for an ACEScg master use our OCIO's
  OCIOLogConvert(logc3) + OCIOColorSpace(Rec.709 -> ACEScg); Gear's edge is LogC4 (a curve our OCIO lacks) plus
  the grade panel.
• Nano Banana 2 Lite - new official recipe (68 -> 69 recipes, 149 -> 150 models). Its own MODELS.md entry:
  the fast / cheap Nano Banana tier, official Comfy partner API nodes / templates api_nano_banana_2_lite_t2i /
  _image_edit (confirmed in the Comfy-Org templates index, not just the launch post). Vendor speed / price
  claims (~4 s, ~$0.034 / 1K) marked as marketing; cloud / paid. Counted as a buildable official recipe AND a
  distinct model per the same rule that counted Gemini Omni Flash.
• Divide and Conquer tiled upscale (Steudio/ComfyUI_Steudio). Auto-tile-sizing upscaler (computes the optimal
  resolution, splits into seamless tiles, merges back) added to ADVANCED.md as an alternative to Ultimate SD
  Upscale / Tiled Diffusion. From the Pixel Pro field-research delta 2026-07-01.

Count change: 68 -> 69 recipes, 149 -> 150 models, from Nano Banana 2 Lite only. LumiPic, ComfyUI_Gear, and the
Divide and Conquer tiler do NOT change the counts - community LoRAs / techniques / tools, on already-counted bases.
All count places synced: README, MODEL_INDEX, and the cover +
models_by_modality banners re-rendered (chip now 150 models · 69 recipes); the models_by_modality chart's stale
per-modality alt text (37/20/66) was corrected to the real 39/21/69 in the same pass.

v2.1.4 - 360 VR panorama, Klein fine-tune, and INT8 acceleration

Added

• 360 / VR equirectangular panorama - Flux.2 Klein AND LTX-2.3 (two separate routes). The shared
  nomadoor/ComfyUI-Panorama-Stickers pack (MIT, Comfy Registry v1.3.0; four ERP nodes: Panorama Stickers /
  Cutout / Preview / Seam Prep) is documented model-agnostically in NODE_LIBRARY/custom-author.md - it is a
  360 projection tool, not a model-specific pack - and referenced from BOTH recipes. (1) Flux.2 Klein 360
  IMAGE (under FLUX.2 in MODELS.md): nomadoor's 360-ERP-outpaint LoRAs (flux-2-klein-4B-... apache-2.0,
  ...-9B-... license:other). (2) LTX-2.3 360 VIDEO (under LTX-2.3): text-to-360 via the public CivitAI LoRA
  360-degree panoramic shot - LTX-2.3 by Ragamuffin20 / Aitrepreneur (civitai.com/models/2327337, direct
  download api/download/models/2816797?fileId=2702793) - the LoRA the Floyo template wraps, which CORRECTS the
  prior "source unconfirmed" note - and flat-to-360 outpaint via TheBurgstall/VR-360-Outpaint-LTX2.3-IC-LoRA
  (cc-by-nc-4.0, v0.1 POC). Full card detail folded in from the CivitAI page: License LTXV2 (not the blanket
  "commercial-OK" I first wrote), trigger A 360-degree panoramic video, weight 0.6-1, aspect 2:1, the wrap-seam
  artifact the author reports fixed (civitai article 25291), plus true-VR finishing steps (mono->stereo via
  SamSeenX/ComfyUI_SSStereoscope, VR metadata via Google spatial-media). The Flux (image) and LTX (video)
  routes are unrelated and kept in their own model sections; neither cross-references the other for the pack.
• Flux2-Klein-9B-True-V3 (community fine-tune) recipe. A wikeeyang aesthetics / composition fine-tune of
  FLUX.2 [Klein] 9B (text-to-image + prompt-only instruct edit + LoRA face-swap / try-on + Mask+LoRA regional
  edit), with its full quant ladder (bf16 / fp8mixed / int8mixedrow / INT8-ConvRot / mxfp8 / nvfp4 /
  GGUF Q4_K..Q8_0). Added under FLUX.2 in MODELS.md, with a license caveat: the card is tagged Apache-2.0 but
  the weights derive from FLUX.2 [Klein], so verify the base license before commercial use. Distinct from
  nomadoor's 360-outpaint Klein LoRA above.
• INT8 inference acceleration note (ADVANCED.md). INT8 weight quantization is now native in ComfyUI
  (~1.5-2x over fp8 on 40-series+ per the source pack), which largely supersedes the ComfyUI-INT8-Fast pack;
  documented when to use the native loader vs the pack, the loader/naming mismatch, and convert_to_comfy.py
  for old quants, using the Flux2-Klein-V3 dual INT8 quants as the worked example.
• OCIO node docs refreshed to ComfyUI-OCIO v1.0.1 (NODE_LIBRARY/ocio.md, + a note in ADVANCED.md).
  Documented the v1.0.1 OCIO Write additions - colorspace_in_name (colorspace before the frame number,
  name_acescg.0086.exr), auto_colorspace (auto Rec.709 -> ACEScg when wired from LTX's HDR decode), and
  Nuke-style EXR compression - plus the new logc3 (ARRI LogC3) curve on OCIOLogConvert and a worked LTX-2
  HDR -> ACEScg EXR-sequence recipe (automatic + manual). Confirmed from the shipped v1.0.1 source; flagged that
  the pack's pyproject.toml still reads 1.0.0 at the v1.0.1 tag (a version-bump miss in the pack, not the docs).

Recipe / model counts are unchanged (still 68 recipes / 149 models). These are community fine-tunes / LoRAs / a
tool built on base models the count already includes (FLUX.2 Klein, LTX-2.3), so they add NO new official model -
the same double-counting reason the (fully PUBLIC) kijai ecosystem is documented in KIJAI.md rather than folded
into the 149. That is a SEPARATE exclusion from private / client-trained models, which is about IP, not
public-vs-not: kijai is fully public. The cover banner does not change.

---

Also included (not a CHANGELOG "Added" entry, tracked in docs/KNOWN_ISSUES.md): the weekly bug-log update - 2 issues fixed, 1 new open, 1 refined.

v2.1.3 - Installer NativeCommandError fix across the per-agent adapters

Patch release: finishes the installer stderr fix started in v2.1.2. Urgent.

Fixed

• The per-agent installers had the same NativeCommandError bug as the shared one. agents/claude/install.ps1 and agents/codex/install.ps1 clone the node-building skills (and codex registers its MCP) with & git clone ... 2>&1 | Out-Null / & codex mcp add ... 2>&1 | Out-Null under $ErrorActionPreference = "Stop", so a harmless native stderr line could abort the adapter install (v2.1.2 fixed only the shared installer). Both now route those calls through the Native helper. The gemini / qwen adapters make no native calls and were unaffected.
• Installer catches a partial template clone instead of reporting success. shared/install_shared.ps1 now gates the "cloned + index built" message on the exit codes of all three git steps (clone / sparse-checkout / checkout), not just a Test-Path index.json check, so a partial clone falls through to the "template clone incomplete" warning.

v2.1.2 - Installer no longer aborts on npm warnings

Patch release: fixes the Windows installer aborting on a harmless npm warning (reported by an installing user).

Fixed

• Windows installer no longer aborts on a non-fatal npm warning. With $ErrorActionPreference = "Stop", PowerShell 5.1 turns a native command's stderr (piped 2>&1 | Out-Null) into a terminating NativeCommandError. So an npm warn deprecated prebuild-install@7.1.3 during npm install -g comfyui-mcp aborted the whole install. shared/install_shared.ps1 now runs npm / git / python through a Native helper that sets $ErrorActionPreference to Continue for the call and gates success on the real exit code (genuine failures still stop the install), and passes npm --loglevel=error.
• Reproduced on PowerShell 5.1 (& ... 1>&2 ... 2>&1 | Out-Null under Stop threw NativeCommandError) and confirmed the fix tolerates the warning while still catching a non-zero exit.

v2.1.1 - Gemini Omni Flash corrected (official GeminiVideoOmni node, buildable)

Corrects the Gemini Omni Flash entry shipped in v2.1.0.

Fixed

• v2.1.0 got the integration wrong. It claimed there was no official Comfy partner node or template for Gemini Omni Flash. Wrong: that leaned on a stale local server (ComfyUI 0.25.1, which does not have the node yet) plus a summarized read of the templates index rather than the raw JSON. A direct read of Comfy-Org/workflow_templates confirms the official partner node GeminiVideoOmni with three shipped templates: api_google_gemini_omni_flash_t2v / _i2v / _video_edit.
• MODELS.md now documents the real buildable graphs: node I/O (model.prompt, model.images.image_1..3, model.videos.video_1..2 -> VIDEO + STRING), T2V / I2V (up to 3 reference images) / conversational video-edit, plus the third-party Anil-matcha/gemini-omni-comfyui (muapi.ai) route as an alternative. MODEL_INDEX updated. Recipe count corrected 67 -> 68, modality chart re-rendered.
• The "replaces Veo entirely" note stands: Veo templates (api_veo2_i2v, api_veo3) are still shipped, so that launch claim remains overstated.

Changed

• Cover banner chip now reads 149 models · 68 recipes (was 68 model recipes) to show the full official model library, not just the hand-written recipes. kijai (62 packs) counted separately; private/client-trained models not in the public number.

v2.1.0 - Gemini Omni Flash documented (cloud and third-party only)

Correction: the Gemini Omni Flash description below is wrong. There IS an official partner node (GeminiVideoOmni) and official templates. Fixed in v2.1.1.

Documents Google's Gemini Omni Flash video model, grounded in primary sources rather than launch copy.

Added

• Gemini Omni Flash (Google). Any-to-any generative video (text-to-video, image-to-video, conversational video editing, native audio; DeepMind model card 2026-05-19). MODELS.md recipe entry + MODEL_INDEX row. A live ComfyUI 0.25.1 /object_info check and the Comfy-Org/workflow_templates index.json confirm there is NO official Comfy Google partner node or local template for it, so the entry says so and points to the real routes (Comfy Cloud, or the third-party Anil-matcha/gemini-omni-comfyui pack via muapi.ai). Recipe count stays 67 (cloud / third-party, not a buildable local recipe).

Changed

• Corrected the "replaces Veo entirely" claim. Google lists Veo and Gemini Omni Flash as separate models, and ComfyUI's Veo 2 / Veo 3 partner nodes are live and not deprecated. Omni Flash is an addition, not a Veo replacement.

v2.0.1 - Self-contained plugin bundle

Patch release: completes what v2.0.0 shipped.

Fixed

• Claude Code plugin bundle is now truly self-contained. tools/build_plugin.py copied only nine core files, so /plugin install comfyui@comfyui-agent-kit shipped without the node library, the node-building guide, and most routed docs, the exact knowledge v2.0.0 added. The builder now mirrors the full installed-skill layout: every routed doc flat, workflow_layout.py, and the whole NODE_LIBRARY/ (incl. ocio.md) as a subdir. Bundle: 9 files -> 17 files + NODE_LIBRARY/ (21 files).

Changed

• README tagline split across two lines: the local-first / every-agent line, then "Your GPU, your models, no cloud, no account."
• Positioning line now leads with experts: "for experts and everyday users".

v2.0.0 - Node library + OCIO color pack + node-building guide

Mega release. The kit grows from a local-first ComfyUI driver into a driver plus a documented node library, a published pro-color node pack, and a field guide to building nodes - local-first, for everyday users and experts alike.

Highlights

• ComfyUI-OCIO companion pack (published v1.0.0): eight Nuke-style OpenColorIO nodes - OCIO Read / Write (load a still / sequence / video, grade in ACES, write ProRes / DNxHR / EXR, fully color-managed, alpha + missing-frame fill + auto frame range over the wire) plus six color operators (ColorSpace, LogConvert, Display, CDLTransform, FileTransform, LookTransform). Documented at buildable depth in docs/NODE_LIBRARY/ocio.md.
• Node knowledge library (docs/NODE_LIBRARY/): 183 per-node entries across every category, plus the full node inventory (552 types across 448 workflows).
• Node-building field guide (docs/BUILDING_NODES.md): the hard-won lessons from shipping ComfyUI-OCIO - widget order, the combo-validation trap, the JS front end, server routes, EXR/ffmpeg traps, verify-on-real-files.
• Workflow auto-layout (shared/comfyui/workflow_layout.py): lay out any graph with zero overlaps, verified in code not screenshots.
• Positioning: local-first, for everyday users and experts.

Full changelog in CHANGELOG.md.