Skip to content

Wan2.2 TI2V 5B

Current status and review date

Decision snapshot

  • Product role: Unavailable text-to-video candidate with functional, capacity, routed, and client evidence but a blocking perceptual failure.
  • Selected or best-qualified configuration: Pinned ComfyUI v0.33.4 workflow video.wan2.2-ti2v-5b-v1, concurrency one, 512×288, 17 frames, eight steps, and 16 fps.
  • Measured hardware: One 32,607 MiB RTX 5090 in an isolated Windows 11, Docker Desktop, and WSL2 media lane.
  • Evidence: A direct 111,055-byte H.264 MP4 completed in 9.092 seconds; a real-Hermes request produced a 117,738-byte MP4 with the same shape. Independent contact-sheet review found severe spatial and prompt- adherence artifacts.
  • Decision: available=false, no-promotion; routed transport works, but the exact workflow failed the bounded quality gate.
  • Important limitation: Generated MP4 binaries were not retained, and temporal smoothness, motion, camera behavior, longer clips, multiple prompts/seeds, and concurrency above one were not tested.
  • Review dates: retained evidence through 2026-09-15; dossier reviewed 2026-09-17.

Review narrative

2026-09-15 — v1 failure and v2 functional repair

The managed v1 functional artifact decoded, but independent review failed all three sampled frames: the requested teapot was unrecognizable amid severe spatial corruption. Official Wan2.2 references exposed v1 graph defects. A new immutable v2 graph restored recognizable red-teapot/table content in all three sampled frames, but review remains partial because steam and handle were missing and artifacts persisted. V2 is functional only, available=false, and no-promotion; this is not a latency comparison.

2026-08-27 — Feasibility screen

The workflow passed the paper and configuration feasibility screen for the isolated RTX 5090 media lane. That result only admitted the candidate to local qualification; it did not establish runtime or output quality.

2026-08-28 — Direct workflow qualification

After correcting the immutable graph for VideoHelperSuite's required save_output input, the pinned workflow produced a decodable 111,055-byte H.264 MP4 in 9.092 seconds. Functional and c1 capacity gates passed. Independent contact-sheet review then failed prompt adherence and spatial quality because severe artifacts dominated the frames.

2026-08-28 — Gateway and Hermes validation

An exact-build cross-host candidate passed gateway transport and a real Hermes request, producing a second decodable 117,738-byte H.264 MP4. This proved the routed/client path, not model quality. The existing production router and unrelated inference services were unchanged.

Immutable identity

Model files

  • Repository: Comfy-Org/Wan_2.2_ComfyUI_Repackaged.
  • Revision: c4f60d30c55a624e35427060fdd217579a6c1d77.
  • Components: diffusion model, UMT5 XXL FP8 scaled encoder, and VAE.

Exact file sizes and SHA-256 identities are recorded in the public workflow bundle lock and the dated evidence.

September v2 workflow

The corrected video.wan2.2-ti2v-5b-v2 graph has SHA-256 b93170eeb578ce5a8dc86d2dfbd944e037abf2e8b9a223059f8e9c6aa8291c14. Runtime reports identify ComfyUI v0.33.4, CUDA 13.0, and PyTorch 2.13.0+cu130; the derived image identity was not independently verified. Small used 20 steps and standard used 30. These settings and outcomes are separate from v1 below.

Runtime and workflow

  • Base container: obeliks/comfyui@sha256:4a40acf4790733dfeb38a92dc2b976c812a484264ff08362e7f8c0d5a16df3e9.
  • ComfyUI revision: 7a131a3afadc8200120f67f9236311a2c48b7445.
  • VideoHelperSuite revision: 4ee72c065db22c9d96c2427954dc69e7b908444b.
  • Workflow: video.wan2.2-ti2v-5b-v1.
  • Graph digest: bd12b2de2a33bbedc91d7ad6120714f3c3adbd174694ac74d6f0213ecef9572e.

Tested hardware and topology

Direct lane

  • One RTX 5090 with 32,607 MiB.
  • Windows 11, Docker Desktop, and WSL2.
  • Isolated media workload at concurrency one.

Routed client lane

  • Exact-build candidate with Primary Node gateway and a real Hermes client.
  • Production router and unrelated inference services remained unchanged.

Other accelerators, multi-GPU execution, and co-resident media generation were not tested.

Engine, quantization, KV, context, and concurrency recipe

Historical v1 functional workflow shape

  • ComfyUI v0.33.4, CUDA 13.0, and PyTorch 2.13.0+cu130.
  • Pinned VideoHelperSuite custom node.
  • Concurrency: one.
  • Resolution: 512×288.
  • Frames: 17.
  • Sampling steps: eight.
  • Frame rate: 16 fps.
  • KV cache and language-model context: Not applicable to this diffusion workflow.

The immutable workflow graph and bundle lock provide the public reconstruction inputs.

Evidence by measurement class

September 15 v2 repair

  • Status: functional; quality remains unverified, no-promotion.
  • Measured: v2 produced decodable 512×288/33-frame and 832×480/49-frame H.264 artifacts. Sampled frames regained recognizable content after v1's severe corruption, but both reviews remained partial.
  • Limits: missing or cropped handles, steam errors, and synthetic artifacts remain. Temporal quality was not assessed; each run is one sample with different parameters, so their timings are not a performance comparison.
  • Evidence: bringup finding and workload/graph identity.

Direct functional and capacity evidence

  • Output: decodable 111,055-byte H.264 MP4.
  • Runtime: 9.092 seconds.
  • Shape: 512×288, 17 frames, 1.0625 seconds.
  • GPU memory: 18,263 MiB peak from a 943 MiB worker baseline.
  • Queue: maximum running one, pending zero.

Routed and client evidence

  • Real-Hermes output: decodable 117,738-byte H.264 MP4.
  • Shape: same resolution, frame count, and duration as the direct run.

Bounded quality evidence

Independent contact-sheet review failed prompt adherence and spatial quality because severe artifacts dominated the frames. The contact sheet could not assess temporal smoothness or camera motion.

Evidence retention

Structured metadata and contact-sheet evidence are retained. The generated MP4 binaries themselves are not retained.

Decision and promotion state

Retained unavailable candidate

  • available=false.
  • no-promotion.
  • Gateway and Hermes transport work for the exact workflow.

Reconsideration gate

A new immutable workflow version plus multi-prompt and temporal review is required before another enablement decision. This record does not authorize a route or deployment change.

Failures and gotchas

Corrected workflow defect

The pinned VideoHelperSuite revision requires save_output. The immutable graph and digest were corrected before the passing run; the earlier graph is not the qualified recipe.

Quality and scope limitations

  • Severe spatial and prompt-adherence artifacts are independently blocking.
  • Temporal smoothness, motion, and camera behavior: Not tested.
  • Longer clips, multiple prompts/seeds, and concurrency above one: Not tested.
  • Generated MP4 binaries: Not retained.
  • One short c1 run does not qualify general throughput.

Dated run history