v3.1.0 - Wan Animate 2, SCAIL-2, FLUX 3 Video and the Seedance 2.5 retraction

Added

• Wan Animate 2 (MODELS/video-open.md), the week's one open-weight arrival. Character animation with no
  pose or keypoint extraction: reference image for identity, driving video for motion, prompt for background
  and camera. Written buildable from comfy_extras/nodes_wan.py on master and the official
  video_wan_animate2.json: the full node chain, WanAnimate2ToVideo's complete input and output list
  including the three strength and window controls, the shipped settings (832x480, 81 frames, cfg 1, 6 steps,
  lcm + simple, shift 5) and why cfg 1 only works with the lightx2v distill LoRA loaded, the
  WanAnimate2Cache speed-for-RAM trade with its standard_static requirement, and the manual subgraph
  chaining needed to pass 81 frames while native loop support is still an unmerged PR.

• FLUX 3 Video (MODELS/video-api.md), BFL's first video model in ComfyUI, with synchronized audio. All
  three nodes, every shared widget, the keyframes and placement mechanics on image to video, and the
  per-second prices read off each class's own rate constants, including the fact that continuation costs
  roughly 2.4x a fresh generation at 720p.

• Seedance 2.5 in ComfyUI (MODELS/video-api.md, seedance skill, seedance/reference.md), and see
  Corrected below for why this is also a retraction. Same ByteDance2* nodes, new model option; the 480p / 720p
  ceiling, mp4-only output, 4 to 30 s duration, the 30 images plus 10 videos plus 10 audios reference capacity
  confirmed in code, the video_editing toggle, the double-quote dialogue rule, and worked prices from the
  node's own badge expression.

• Seedream 5.0 Pro Layer Separation (MODELS/image-api.md): one image into a background plate plus up to
  16 transparent layers, the <bbox> per-mille targeting syntax, all six outputs in slot order, both recompose
  paths the official template wires, the inverted-mask gotcha, and the price.

• Topaz Bloom 2 and Wonder 3.5 (MODELS/utility.md) on TopazImageEnhanceV2, with the widgets each model
  reveals, Wonder 3.5's 1x to 6x limit, and the aspect-ratio behaviour that stops you using either to reframe.

• Qwen Image 3.0 Pro (MODELS/image-api.md) as an explicit dead end: the templates are gated to the
  Comfy Cloud distribution and the node classes exist in no public source. The entry exists so nobody
  spends an afternoon hunting for nodes that are not there.

• SCAIL-2 promoted from a four-line note to a full buildable entry and moved from MODELS/niche.md to
  MODELS/video-open.md, where the other local video models live. The graph as the official template builds
  it: SAM3_VideoTrack twice into SCAIL2ColoredMask into WanSCAILToVideo, the six-colour palette that is
  the character binding, sort_by as the thing that keeps an identity on one colour across both outputs, the
  replacement_mode flag that must agree in both nodes, the chunked long-video path, and the shipped
  settings including a second DPO LoRA that no write-up mentions and a ColorTransfer pass that matches
  the generated character back to the plate.

Corrected

• The kit sent you to a community pack for SCAIL-2 when ComfyUI has had it in core since 2026-06-09. PRs
  14373 and 14509, nodes in comfy_extras/nodes_scail.py, two official templates, detection by state-dict key
  so a plain UNETLoader finds the checkpoint. Two smaller errors went with it: pose_strength was described
  as "exact-copy vs style adaptation" when it is a latent multiplier that does not exist upstream at all, and
  the mask preprocessor was named as a bundled SCAIL-Pose when the native path uses SAM3.
• Reference-to-video on MiniMax H3 has a six-section output contract the skill did not know. MiniMax ships
  two prompt guides; the kit had only the base one, whose three fields are correct for T2VA / I2VA / FL2VA /
  L2VA. Full-reference mode takes subject_definitions, summary, retention_analysis,
  detailed_description, overall_soundscape, non_diegetic_music, in that order. Since the core node
  validates none of it, nothing reports the mistake except the result.
• A text encoder that does not exist was listed for Wan Animate 2. The Comfy-Org card advertises
  umt5_xxl_fp16; the repository does not contain it. Copied from the card without checking the file listing.
• Four capabilities were being read as available on Wan Animate 2 that ComfyUI cannot reach:
  multi-character, text-driven viewpoint control (the "Viewpoint LoRA" has no published weights anywhere,
  checked five ways), the streaming Lite variant (unreleased), and character replacement, which is a
  regression against Wan Animate v1 rather than a feature. Named explicitly so the gap stops reading as
  silence.
• "Seedance 2.5 has no ComfyUI nodes" was true on 2026-08-01 and wrong from 2026-08-08. Core v0.31.0 shipped
  it. The claim appeared in three places (MODELS/video-api.md, seedance/SKILL.md, seedance/reference.md)
  and every one is corrected in place with the old claim named, not silently overwritten. It is the second time
  in three weeks that a dated "checked" line went stale inside a week.
• Reve is deprecated, not merely quiet: all three nodes carry is_deprecated=True in core v0.31.0 and the
  four templates were deleted from the official library. Marked in MODELS/image-api.md, MODEL_INDEX.md and
  the bug log.

Changed

• docs/KNOWN_ISSUES.md rebuilt for v0.30.0 and v0.31.0. The headline is a MiniMax H3 bug cluster: ten
  H3 issues filed from the v0.31.0 release date onward, five of them now rows here (noise output, sampler RuntimeError,
  17-frame dark-frame artifacts, an unrecoverable VAE decode OOM whose tiled fallback is a no-op, and a roughly
  2x slowdown), against five H3 fixes that the same release delivered. Also added: Kling legacy model and
  Virtual Try-On removal, Reve deprecation, ImageUpscaleWithModel on 4 GB cards, and a Load Checkpoint memory
  leak. The Custom Combo subgraph row was not promoted to "Recently fixed" despite being closed as
  completed, because the last comment on the issue says it still reproduces.
• Six defects found by an audit of the kit against itself, all fixed: SKILL.md still claimed 156 models
  after the rest of the repo moved to 160; README listed Wan variants as a bare "Animate", which stopped being
  unambiguous the day Animate 2 shipped; minimax-h3/reference.md had two escape sequences collapsed into real
  control characters, so models\text_encoders\ and models\vae\ rendered as garbage and sent files to the
  wrong folders; the seedance skill existed but was absent from the routing map, so an agent reading only
  SKILL.md never learned of it; the precision ladder in SKILL.md omitted int8-convrot entirely while
  ADVANCED.md calls it half the size of fp16, faster, and a match for fp8, with Ampere supported; and the H3
  weight table was missing the 40.23 GB pruned_bf16 row.
• New in the H3 skill: Sol-Attn, both packs, with the correct finding that the widely repeated "Blackwell
  SM120 only" is a README line and not a code gate, so it loads and takes the sparse path on Ampere, plus a
  real defect in one pack's fallback chain. And the tiny-autoencoder preview path, which is 10 MB, never
  touches VAEDecode, and rides on ModelPreviewOverrideKJ's tiny_vae slot.
• Library counts resynced after a week that removed more templates than it added: 586 to 574 templates
  (the Getting Started category was archived and the Reve and Hailuo MiniMax templates were deleted) and
  156 to 160 distinct models. Recipe entries 68 to 72. Updated in README, MODEL_INDEX.md,
  MODELS.md's router table, the three chart generators, cover_gen.py, and the four rendered banners.