Wan Animate continuation LoRA — training and evaluation

Purpose

Improve appearance retention during chained Wan Animate dance generation. This experimental adapter is not a universal color correction or anatomy repair model.

Released checkpoint

File: wan_cuda_expanded1000_20261012.safetensors
SHA256: f62e722c6e05daedc01906149a6f5c99d6fc7f1e4a3991b5749ce46c7f40f5bb
Base family: Wan2.2-Animate-14B; evaluated with the Orca INT8ConvRot inference checkpoint.
Rank: 16. Trainable image-conditioning K/V adapters plus an inherited frozen self-attention Q/V adapter (parent contribution 0.5 baked into the checkpoint).
Recommended tested adapter strength: 1.0. Disable the trained adapter for a baseline comparison.
LightX2V loader is retained at strength 0 in these comparisons. It is a separate adapter.

Training process

Initial small pilot used actual five-frame continuation conditioning. First-frame-only training was insufficient for the production continuation task.
Expanded qualification selected 14 clean early windows from approved generated videos: 11 training windows and 3 held-out windows. The acquisition range was frames 7–90; the trainer consumed 65-frame windows, frames 7–71: five conditioning frames (7–11) and 60 target frames (12–71). Dreamin Chuchu and Mio POP IN 2 were priority sources.
BF16 backbone training used a target-only latent flow loss. Reference/context positions were excluded from the loss; the first three latent positions were masked for this implementation's temporal packing.
The released CUDA run warmed from checkpoint 260, trained on the expanded dataset to step 1000 at learning rate 5e-6. Fresh image-only and joint-adapter runs, component isolation, precision controls and checkpoint comparisons were evaluated separately.
Evaluation included five recursive five-frame generations, native-resolution comparisons, normal-length clips and unseen characters. Merely loading a zero-strength adapter, switching precision, or adding clean-hair prompt wording did not reliably eliminate texture defects.

Reproducible unseen evaluation protocol

Two new synthetic adult female references, absent from training: a realistic brown-haired person and an original blonde anime character. Both use a daylight seaside stage and held-out Melancholic choreography. Both primary A/B tests ran on the same RTX 5090 hardware, 576×768, six Euler/simple steps, CFG 1, sampling shift 5, LightX2V strength 0. No Qwen editing or color postprocessing was applied.

Both continuation arms start from the same five baseline-generated usable opening frames. Each round generates five new frames; those become the next round's context, repeated five times. A separate normal-length comparison generates 56 new frames from the same starting context. Matching seeds, pose controls and references are used within each pair. Immutable opening frame 0 is excluded. All 28 primary prompts succeeded.

Results and limitations

The trained adapter retains sharper detail and reduces haze after repeated continuation in both unseen cases. It also darkens the whole scene and increases saturation more than baseline. Head-region color proxies are closer after five rounds, but normal-length baseline color is generally closer. Added dark hair lines and arm/skin texture noise remain. Darker hair from added lines must not be counted as a successful hair-color correction.

These are two synthetic characters, one choreography and one seed family. Color masks are spatial/color proxies, not semantic ground truth or identity/anatomy scores. No claim of universal generalization is supported.

Use and compatibility

In the deployed Orca dance recipe v5, trained_lora_enabled defaults to true; pass false to disable. Existing compiled recipes remain pinned. For another runtime, load the adapter on the matching Wan Animate model and validate key coverage, numerical precision and controls before relying on the result. The API flag is Orca-specific, not a standard ComfyUI field.

Image comparisons

Baseline means no trained continuation adapter. Both arms retain LightX2V at strength 0. Each round generates five new frames using the preceding five frames as context. All examples are native images without color postprocessing.