v2.5.0 - HeyGen avatar stack and JoyAI Image Edit

[2.5.0] - 2026-07-25

Added

• HeyGen - new official recipe (avatar video, talking photo, video translate, TTS). A presenter/avatar stack,
  not a scene generator, and the five nodes each do one job: HeyGenTalkingPhotoNode (LoadImage -> image,
  VIDEO out), HeyGenAvatarVideoNode (standalone, VIDEO out), HeyGenCreateAvatarNode (avatar_id STRING
  • preview IMAGE out; feed the ID into Avatar Video's custom_avatar_id), HeyGenVideoTranslateNode
    (LoadVideo -> video, VIDEO out) and HeyGenTextToSpeechNode (AUDIO out). Documented the two DynamicCombo
    widgets that decide what you can even see: speech (script exposes text / voice / custom_voice_id /
    voice_speed; audio replaces them with an AUDIO input) and engine on Avatar Video (auto / avatar_iv /
    avatar_iii / avatar_v, each filtering the avatar list, with the per-second price of each). Also the traps: a
    voice is required on Talking Photo but optional on Avatar Video, background_color must carry a leading #, the
    2000px auto-downscale, the 5000-character script cap, and seed being a re-run trigger that HeyGen never
    receives. Confirmed from comfy_api_nodes/nodes_heygen.py on master plus templates
    api_heygen_{avatar_video,talking_photo,text_to_speech,video_translate}.
• JoyAI Image Edit (JD, Apache-2.0) - new official recipe. Native core support via
  comfy_extras/nodes_joyimage.py: one node, TextEncodeJoyImageEdit, which tokenizes the instruction WITH
  the reference images and appends their reference_latents when a VAE is connected. Full graph documented
  (CLIPLoader at type joyimage, the Wan 2.1 VAE, ImageScaleToTotalPixels at 1 MP feeding both the
  node and GetImageSize -> EmptySD3LatentImage, CFGNorm on the MODEL line rather than the conditioning,
  40 steps / CFG 4.0 / euler / normal), plus the images Autogrow socket (images.image0, zero-based and
  namespaced, up to six slots) for multi-reference edits and the 49 fixed ~1MP aspect buckets. Flagged a broken source: the Comfy-Org/JoyAI-Image-Edit card body was copy-pasted from the
  Plus release and lists joyai_image_edit_plus_* filenames the repo does not host; the real files were
  confirmed by listing the repo.
• Anima ControlNet-LLLite - control and inpainting for the existing Anima recipe. Loads as a MODEL_PATCH,
  so ControlNetApply is the wrong node entirely: ModelPatchLoader (from models/model_patches/) ->
  AnimaLLLiteApply (model + model_patch + image control map + optional mask, knobs strength /
  start_percent / end_percent) -> sampler. Documented all ten published patch files (any / depth / lineart /
  pose / scribble / inpainting), the rule that a mask only applies to a 4-channel inpainting patch and is silently
  discarded otherwise, the Depth Anything 3 chain that builds the depth map
  (LoadDA3Model -> DA3Inference -> DA3Render), and the template settings. Licence flag: non-commercial.
• Krea 2 Turbo image style reference on CORE nodes. The krea2_style_reference LoRA (ostris) generates in the
  style of a reference image with no trigger word. The HF card says a custom node is required; the official
  template image_krea2_turbo_int8_image_style_reference does it with core nodes only, and that full chain is now
  documented (LoraLoaderModelOnly -> ModelSamplingFlux -> CFGGuider at cfg 1.0,
  TextEncodeQwenImageEditPlus carrying the reference, FluxKontextMultiReferenceLatentMethod at
  index_timestep_zero, 8 steps). Licence flag: krea-2-community-license.

Changed

• Credits and licensing brought up to date. ATTRIBUTION.md and the README "Credits and thanks" now name
  kohya-ss for Anima ControlNet-LLLite (non-commercial), ostris for the Krea 2 style-reference LoRA, and JD
  (jd-opensource) for JoyAI Image Edit. Two pre-existing gaps were closed at the same time: ostris's
  ComfyUI-Krea2-Ostris-Edit nodes and reverentelusarca's detail-enhancer LoRA had been documented in
  MODELS.md since an earlier release with no attribution row.
• Counts refreshed for the 2026-07-25 template sync: 74 recipes (was 72), 157 indexed models (was 153),
  578 templates (was 562, +16 this week); 94 Subgraph Blueprints unchanged. README, docs/MODEL_INDEX.md, the
  GitHub About box, the three chart generators and tools/assets/cover_gen.py were all updated and the four
  banner PNGs re-rendered and visually verified.
• docs/KNOWN_ISSUES.md refreshed to 2026-07-25. Four new open bugs: the core commit 7c59a078d that removed
  interleaved_freqs_cis and kills the whole ComfyUI-LTXVideo pack on master, int8_convrot NaN/black output on
  RDNA4 ROCm, Qwen Image Edit silently exiting on PyTorch cu130, and Custom Combos not updating their index output
  inside subgraphs (which breaks Comfy's own Blueprints). Nothing moved to "Recently fixed": four of the five
  closures in the window were NOT_PLANNED, not fixes. Also recorded two maintenance gotchas: the
  comfyanonymous/ComfyUI -> Comfy-Org/ComfyUI redirect that the GitHub search API refuses to follow, and the
  need to read stateReason before calling a closed issue fixed.