Wan2.1 SCAIL-preview ComfyUI: reference/pose routing, half-size poses and alignment controls

Wan2.1 SCAIL-preview pose control: reference routing, half-size poses and alignment This is source-only analysis of kijai/ComfyUI-WanVideoWrapper's example_workflows/wanvideo_2_1_14B_SCAIL_pose_control_example_01.json at commit 4b709a7a04e67b2332dad017b5eee18514c68182. The pinned graph's raw SHA256 matches the source_file_sha256 reported by the retained discovery record. Its different workflow/canonical digest fields are preserved separately. This does not establish what those normalization algorithms did, a successful render or independent verification. No workflow, model, loader or installation was executed, and no separate reference image, driving video, weights or generated output was fetched. The original JSON includes a cached scheduler plot and preview metadata; those remain original source data, not our execution evidence. This example uses a reference image and a driving video. It is a Wan2.1 14B SCAIL-preview custom-wrapper graph, distinct from the separately documented SCAIL-2 model. Reference and pose-conditioning nodes do not establish character retention across a camera-angle change. The useful source-backed result here is the exact input/routing and alignment recipe, with that retention question still open. Follow the named connections as well as the visible wires. The saved graph has one setter for each of nine names and twelve corresponding getters. For example, cfg is FloatConstant238 -> SetNode239 -> GetNode240 -> sampler348; VAE is loader38 -> setter173 -> getters174/176. These are serialized name-based routes, not additional visible LiteGraph links. A runtime/frontend capable of those utility nodes is still required and was not tested here. Reference path: LoadImage106 -> ImageResizeKJv2 node107 -> reroute351. The resized image goes both to WanVideoAddSCAILReferenceEmbeds350 and to WanVideoClipVisionEncode327, whose output also enters350. The CLIP-vision loader326 selects clip_vision_h.safetensors. Empty image embeddings99 enter350 first; reference embeddings350 then enter pose embeddings324; the resulting image_embeds go to WanVideoSamplerv2 node348. The VAE getter174 supplies both embedding nodes. This is the connected reference path in this example; it does not verify the discovery summary's broader suggestion that reference inputs can be omitted without changing the graph. Driving/pose path: VHS_LoadVideo130 -> ImageResizeKJv2 node131 -> reroute370. That frame batch feeds NLFPredict334 using model loader335, and driving pose detection363 using ONNX model loader364. Reference pose detection365 receives the resized reference image107 using the same detection-model loader. RenderNLFPoses362 receives NLF results334, driving dw_poses363 and reference ref_dw_pose365. Its rendered pose image passes through reroute338 into WanVideoAddSCAILPoseEmbeds324. The pose-render preview branch is separate from the sampled output path. The important sizing controls are linked. Width constant203 stores512 and height constant204 stores896, routed through named width/height setters206/207 and getters208/209 to both image resizers. They override reference-resizer107's saved480x832 widgets. Generation dimensions then come from the reference resizer's width/height outputs through gen_width/gen_height setters184/185 and getters186/187 into empty embeddings99. Do not copy99's saved480x1216 as the effective connected dimensions. Pose rendering has its own half-size path: width/height getters341/342 -> calculators371/372, both storing a / 2 -> RenderNLFPoses362 width/height. For the declared512x896 controls,256x448 is the arithmetic target, overriding that renderer's saved512x896 widgets. The embedded author note says pose resolution should be half the generation resolution; another note says dimensions should be divisible by32. This is a documentary consistency check, not measured output geometry or a rule verified for every SCAIL model/version. Drawing and alignment are different controls. The embedded note says optional dw_poses supply face/hand drawing and alignment to the reference image. It says alignment still occurs when drawing is disabled while dw_poses remain connected, and instructs disconnecting them to disable alignment. Both driving dw_poses and ref_dw_pose are connected in this saved graph. The exact effect of changing each connection was not executed; do not infer that a drawing toggle alone removes reference alignment or guarantees identity consistency. Frame count is connected from VHS_LoadVideo130's frame_count output through gen_frame_count setter375/getter377 into embeddings99. The loader stores frame_load_cap81, select_every_nth2, skip_first_frames0 and force_rate0. Neither that cap,99's saved65-frame widget nor resizer131's cached161-frame display is an observed frame count here. The main output node139 stores16 FPS and has no connected audio. A saved preview frame rate is cache metadata and need not reflect the current widget; these values do not prove that input timing, selected frames and output timing match. The source declares Wan21-14B-SCAIL-preview_fp8_e4m3fn_scaled_KJ.safetensors, a Wan2.1 VAE, umt5-xxl-enc-bf16.safetensors, clip_vision_h.safetensors, a Lightx2v I2V distillation LoRA, ViTPose and YOLO files, plus an NLF TorchScript URL. The model filename contains SCAIL-preview; do not replace it with a SCAIL-2 checkpoint or apply SCAIL-2's advertised capabilities without separate compatibility evidence. The embedded model note explicitly says to use the I2V distillation LoRA. These are source-declared references, not downloaded, permission-cleared or runtime-compatible components verified by this review. The NLF loader's download-looking node is not an instruction this review executed. Sampler348 is connected to scheduler349, the final reference-plus-pose embeddings, named model/text routes and named CFG. CFG is wired from constant1; the sampler stores seed42 in fixed mode. Scheduler349 stores [dpm++_sde,6,7,0,-1] before its cached plot, while the sampler also has its own saved numeric widgets. Without the matching node implementation/runtime, this review does not equate those positional fields with a tested step/shift contract. The block-swap, compilation and context-option nodes likewise describe a stored configuration, not a proven VRAM fit or speed claim. For a real trial, supply permitted reference/driving media and record the exact graph, model and custom-node versions, input hashes, prompt, selected input frames, size/alignment connections, scheduler/CFG/noise settings and output FPS. Evaluate pose following separately from face, costume and identity retention, including occlusion and changed viewpoints; that evaluation is proposed and unperformed here. The graph mixes saved extension version labels and is not a tested environment lockfile or minimum-runtime specification. The repository's same-commit full LICENSE declares Apache2.0 and is retained with unmodified source bytes and embedded attribution. A separate upstream NOTICE file was not supplied with this capture and has not been checked by this review. The generated provenance memo is not an upstream NOTICE. Model weights, pose dependencies, other custom nodes, source media, likenesses and outputs retain their separate terms. No Hivemind community body, cached plot or source media is reproduced in this explanation.

{
  "citations": [
    {
      "url": "https://raw.githubusercontent.com/kijai/ComfyUI-WanVideoWrapper/4b709a7a04e67b2332dad017b5eee18514c68182/example_workflows/wanvideo_2_1_14B_SCAIL_pose_control_example_01.json",
      "note": "Pinned original document; exact byte ranges and hashes are preserved in the curation source pins."
    },
    {
      "url": "https://raw.githubusercontent.com/kijai/ComfyUI-WanVideoWrapper/4b709a7a04e67b2332dad017b5eee18514c68182/LICENSE",
      "note": "Pinned original document; exact byte ranges and hashes are preserved in the curation source pins."
    }
  ]
}