video-extend

Video Extension (Pusa 2.2 — temporal flowmatching)

Overview

Pusa extends a video temporally — it continues / lengthens an existing clip rather than regenerating it from scratch. It does this on the ComfyUI-WanVideoWrapper stack (kijai) using the WAN 2.2 T2V A14B dual HIGH/LOW models you already have for wan-t2v-video, plus the small Pusa V1 LoRAs and a Pusa-specific sampling path: the flowmatch_pusa scheduler and the WanVideoAddPusaNoise node. The input clip is encoded with WanVideoEncode and injected as the first latents of the generation — that is what carries the existing motion/content into the continuation.

The official reference graph is kijai's wanvideo_2_2_14B_Pusa_extension_example_01.json (in ComfyUI-WanVideoWrapper/example_workflows/). This skill is built directly from that workflow plus the live node schemas.

Relationship to wan-t2v-video: Pusa rides on the exact same WanVideoWrapper stack — same T2V A14B HIGH/LOW fp8 models, same UMT5 text encoder, same WAN VAE, same block-swap/torch-compile machinery. The only new downloads are the two Pusa V1 LoRAs (~1.9 GB total). Read wan-t2v-video first for the base stack; this skill is the temporal-extension delta on top of it.

⚠️ Verification note: every node, model, LoRA filename and setting below was confirmed against the live ComfyUI /object_info (WanVideoWrapper installed) and against kijai's example workflow JSON + HF repo (June 2026). Where a value is a starting recommendation rather than a hard requirement it's flagged. Don't substitute a node you can't confirm with list_installed_nodes / get_node_info.

---

What "temporal flowmatching" means here (why it extends, not regenerates)

WAN is a flow-matching video model: sampling integrates a velocity field from noise to a clean latent, and every frame normally shares the same denoising timestep. Pusa's contribution (Vectorized Timestep Adaptation) is to make the timestep per-frame: the frames you already have can be held at (or near) t = 0 (clean) while the new frames start from t = 1 (noise), and the model flow-matches the noisy tail conditioned on the clean head.

Concretely in the graph:

1. WanVideoEncode turns the tail of your loaded clip into a clean latent.
2. That latent is placed at the front of an otherwise-empty embed (WanVideoEmptyEmbeds + WanVideoAddExtraLatent), so the generation's first latents are your real footage.
3. WanVideoAddPusaNoise assigns small, ramping per-latent noise multipliers to those conditioning latents (so they stay mostly clean) and full noise to the new latents — this per-frame noise schedule is the "vectorized timestep."
4. flowmatch_pusa on WanVideoSampler integrates that mixed-timestep field.

Because the conditioning latents are real (not just a single start image like I2V), the continuation inherits the existing motion, subject, camera and color, then keeps going. That's the difference from plain T2V (no memory of any clip) and from I2V (conditions on one still frame only).

---

⭐ Recommended pipeline (the kijai extension graph)

VHS_LoadVideo (your clip)
      │ IMAGE (all frames)
      ▼
ImageResizeKJv2  ◄── resize to 832×480 (divisible by 16), get W/H
      │
      ├─► GetImageRangeFromBatch (tail N frames) ─► WanVideoEncode (vae, image)
      │                                                   │ LATENT  = clean
      │                                                   ▼   conditioning latents
      │                                          GetLatentSizeAndCount ─► count
      │                                                   │
WanVideoEmptyEmbeds (W,H, total_frames=81)                ▼
      │ WANVIDIMAGE_EMBEDS                       CreateScheduleFloatList
      └────────► WanVideoAddExtraLatent ◄────────┘ (per-latent noise multipliers,
                       │  (encoded clip latent at front)   ramp e.g. 0→0.2)
                       ▼ WANVIDIMAGE_EMBEDS
              WanVideoAddPusaNoise  ◄── noise_multipliers (list), noisy_steps
                       │
        ┌──────────────┴───────────────┐
        ▼ (pass 1, HIGH)               ▼ (pass 2, LOW)
 WanVideoSampler (HIGH model           WanVideoSampler (LOW model
   + Pusa HIGH LoRA + distill,           + Pusa LOW LoRA + distill,
   flowmatch_pusa, steps 6, cfg 1,       flowmatch_pusa, steps 6, cfg 1,
   shift 5, start 0 / end 3)             shift 5, start 3 / end -1)
        └──────────────┬───────────────┘
                       ▼ LATENT
                 WanVideoDecode (WAN VAE)
                       │ IMAGE
                       ▼
                 VHS_VideoCombine  ─► MP4 (16 fps)

• VHS_LoadVideo / VHS_VideoCombine come from ComfyUI-VideoHelperSuite (installed). VHS_VideoCombine is preferred for the encode (audio passthrough).
• Everything WanVideo* is ComfyUI-WanVideoWrapper (installed).
• ImageResizeKJv2, GetImageRangeFromBatch, GetLatentSizeAndCount, CreateScheduleFloatList are ComfyUI-KJNodes (installed alongside the wrapper). They're convenience nodes — see "Minimal wiring" if you want fewer.

The two load-bearing nodes (confirmed schemas)

WanVideoAddPusaNoise — "Adds latent and timestep noise multipliers when using flowmatch_pusa."

It outputs WANVIDIMAGE_EMBEDS straight into WanVideoSampler's image_embeds.

flowmatch_pusa — a value in WanVideoSampler.scheduler (confirmed present in the dropdown: ...flowmatch_distill, flowmatch_pusa, multitalk...). It must be selected on the sampler(s) for the Pusa noise schedule to be interpreted correctly. The example also wires explicit WanVideoScheduler nodes set to flowmatch_pusa, steps 6, shift 5 (one per pass, split 0–3 / 3–end).

How the input clip conditions the extension (the key wire)

WanVideoEncode(vae, image=<tail frames of clip>) → LATENT → WanVideoAddExtraLatent (or WanVideoEmptyEmbeds.extra_latents, tooltip: "First latent to use for the Pusa -model"). This places the real clip's latents at the head of the embed window. The sampler then only has to generate the tail, flow-matched onto that clean head — that is the entire trick. No CLIPVision, no WanFirstLastFrameToVideo.

---

In practice: load → strip → re-point (DON'T hand-build) ⭐ preferred

The kijai wanvideo_2_2_14B_Pusa_extension_example_01.json is a 56-node graph thick with GetNode/SetNode buses, Reroutes, and an alternate (dead) text branch. Hand-wiring the Pusa noise / extra-latent / frame-stitch path is slow and error-prone. The reliable flow is load the real graph, then adapt ~7 widgets:

1. Stage the example anywhere on disk (e.g. copy into the ComfyUI workflows folder).
2. panel_load_workflow(path: …) — drops it on the canvas server-side (no 150KB JSON through chat).
3. panel_strip_workflow(path: …) — returns the resolved API graph (Get/Set/Reroute/bypass collapsed to real links). This is how you SEE what is actually wired — it exposes both the dead text branch and the silently-reset dropdowns below. (Raw UI JSON hides them.)

⚠️ TRAP 1 — the example's model paths reset to the WRONG file on load

The example references models by subfolder (WanVideo\2_2\…, WanVideo\Lightx2v\…, wanvideo\Wan2_1_VAE_bf16…). On a flat local models/ layout those don't resolve, so ComfyUI silently falls each dropdown back to the first entry in the list — e.g. both WanVideoModelLoaders land on Qwen_Image_Edit-Q8_0.gguf and the WanVideoVAELoader on LTX23_audio_vae_bf16. It looks wired but errors (wrong arch) or renders garbage. After loading, set each explicitly:

The official HIGH-underscore / LOW-dash filename inconsistency is a real trap — verify each one rather than copy-pasting.

⚠️ TRAP 2 — the distill LoRA silently drops to none

The example's lightx2v path is WanVideo\Lightx2v\…rank64_bf16_.safetensors (note the trailing _). Locally you usually have rank128 (…rank128_bf16), so the slot resets to none on load — which removes the speed LoRA, and 6-step / cfg-1 sampling then produces mush. Re-add it to lora_1 (strength 1.0) on both WanVideoLoraSelectMulti nodes. Keep merge_loras=false on both (fp8 gotcha above).

⚠️ TRAP 3 — the active prompt is on WanVideoTextEncodeCached, not CLIPTextEncode

The example also contains a CLIPLoader → CLIPTextEncode → WanVideoTextEmbedBridge branch (the "red panda" prompt). It is NOT wired to the samplers — both WanVideoSampler.text_embeds come from WanVideoTextEncodeCached (umt5-xxl-enc-bf16). Edit the prompt THERE; the CLIPTextEncode pair is a decoy that strip_workflow will show dangling.

⚠️ TRAP 4 — match the conditioning fps to WAN-native (16)

If your source clip was frame-interpolated (e.g. RIFE'd to 32/50 fps), set VHS_LoadVideo.force_rate = 16 so the conditioning frames carry motion at WAN's native cadence. Otherwise the encoded "past" runs at 2–3× the model's pace and you get a velocity jump at the seam — the exact artifact Pusa exists to avoid. Best practice: extend the pre-interpolation 16fps master, then interpolate/upscale the combined result afterwards, not before.

⚠️ TRAP 5 — the example assumes SageAttention + torch.compile (triton)

WanVideoModelLoader in the example sets attention_mode: sageattn and wires a WanVideoTorchCompileSettings (inductor) into compile_args. Both are optional accelerators with extra deps that a stock Windows ComfyUI usually lacks:

• sageattn → needs the sageattention package. Missing → the model loader hard-fails with ValueError: Can't import SageAttention: No module named 'sageattention' before any sampling. Fix: set attention_mode → sdpa on both WanVideoModelLoaders (always available; a bit slower).
• inductor torch.compile → needs triton (no official Windows build). Missing → compile errors later. Fix: disconnect WanVideoTorchCompileSettings from each model loader's compile_args (or don't load it). Only re-enable these two if you've actually installed sageattention / triton-windows.

Check first with the ComfyUI startup log (it prints Could not load sageattention… and triton: unavailable) or list_installed_nodes.

Preferred end-to-end order

generate (or Krea2→WAN/LTX i2v) → Pusa-extend at 832×480/16fps → THEN upscale+interpolate (hand the extended clip to the video-upscale block / a saved Upscale4x-RIFE-1080p subgraph). Upscaling/interpolating before extending wastes the work and feeds Pusa an off-cadence, harder-to-match conditioning clip.

---

Models, LoRAs & where to get them

UNET — WAN 2.2 T2V A14B (already installed for wan-t2v-video)

Text encoder + VAE: same as wan-t2v-video — UMT5 (umt5_xxl_fp8_e4m3fn_scaled / umt5_xxl_fp16) via the wrapper's text-embed path, and the WAN VAE (wan_2.1_vae) via WanVideoVAELoader. The example uses WanVideoTinyVAELoader + taew2_1.safetensors for fast preview decode; use the full WAN VAE for final-quality decode.

Pusa V1 LoRAs — the ONLY new download (~1.9 GB)

From kijai's HF repo Kijai/WanVideo_comfy, folder Pusa/ → place in models/loras/ (the example expects them under loras/WanVideo/Pusa/):

There is also a single-file Wan21_PusaV1_LoRA_14B_rank512_bf16.safetensors (~4.9 GB) in the same folder — that's the Wan 2.1 single-model Pusa LoRA. For the 2.2 dual HIGH/LOW extension graph, use the two Wan22_...rank_98 files above, matched to the correct expert. Upstream weights / paper: RaphaelLiu/PusaV1 on HF.

Speed LoRA (paired with Pusa in the example)

The example also stacks the lightx2v T2V distill LoRA on each model via WanVideoLoraSelectMulti, so 6-step low-CFG sampling works:

LoRAs are selected with WanVideoLoraSelectMulti (multi-slot) and fed into each WanVideoModelLoader's lora input — one select feeds HIGH (Pusa HIGH + distill), one feeds LOW (Pusa LOW + distill).

⚠️ CRITICAL — merge_loras=false on fp8 models (same gotcha as wan-t2v-video)

Pusa loads LoRAs onto the fp8-quantized T2V A14B models (quantization=fp8_e4m3fn_scaled). As documented in wan-t2v-video: when a LoRA is applied to an fp8 model via the wrapper's LoRA select, set merge_loras to false. The default merge_loras=true tries to bake the LoRA into the already-quantized fp8 weights and hard-crashes ComfyUI during LoRA loading with no Python traceback (looks like an unexplained restart/OOM). false applies the LoRA as a runtime patch, which is fp8-safe. This applies to both the Pusa LoRAs and the lightx2v distill LoRA. Use merge_loras=true only on non-quantized bf16/fp16 models.

---

Settings

Sampler (from the example — distilled 6-step, two-pass HIGH→LOW)

If you drop the distill LoRA: use steps ~20–30, cfg ~5–6, keep flowmatch_pusa and shift 5, single-pass unipc-style splitting still works HIGH→LOW.

Pusa noise (WanVideoAddPusaNoise.noise_multipliers)

This is the dial that controls how strictly the continuation honors the input clip vs. how free it is to diverge:

• Lower multipliers (→ 0) = conditioning latents stay clean = the continuation clings tightly to the source frames (less drift, but can look "stuck"/repeat).
• Higher multipliers = more noise on the conditioning latents = the model is freer to evolve the scene (more new motion, more drift risk).
• The example ramps them [0.0 … 0.2] across the conditioning latents (one per encoded latent, via CreateScheduleFloatList driven by GetLatentSizeAndCount) so the oldest frame is locked and the seam frame gets a little noise for a smooth blend. Start there; nudge the top of the ramp up (~0.3) if continuations feel frozen, down if they drift.

Seam color/saturation drift → ColorMatch the generated frames ⭐

The most common quality complaint with a Pusa extension: the moment you cross the seam, the color saturates / shifts. The conditioning frames are your real footage (near-clean latents), but the generated tail comes purely from the model's prior — which biases toward higher contrast/saturation (worse with the distill LoRA and fp16_fast). Motion carries fine; the palette pops.

Two fixes, best applied together:

1. base_precision: bf16 on both WanVideoModelLoaders instead of fp16_fast. fp16_fast's reduced precision drifts over the generated tail and compounds the saturation; bf16 is more color-stable (small speed cost).

2. Re-grade the generated frames to the source palette with a ColorMatchV2 (KJNodes) between WanVideoDecode and the final stitch/save:

   • image_target ← WanVideoDecode (the generated window)
   • image_ref ← the resized original clip (ImageResizeKJv2 output — your real footage)
   • method: hm-mkl-hm (histogram→MKL→histogram; strongest at removing a palette jump while keeping per-frame variation), strength 1.0.
   • Re-route the downstream consumers (ImageBatchMulti / ImageConcatMulti's image_1) to take the ColorMatch output instead of the raw decode.

   Tune: if under-corrected, raise strength; if washed/over-corrected, drop to ~0.6; for an even tighter temporal lock use a single clean reference frame (the last conditioning frame) instead of the whole clip. Use ColorMatchV2 (not the deprecated ColorMatch).

This also matters for chaining — color-match every new segment to the previous one before concat or the drift compounds hop-to-hop.

Length, frame counts & fps

• WanVideoEmptyEmbeds.num_frames is the total window (conditioning frames + new frames). The example uses 81 total (the WAN-native 4n+1 length, ~5 s @16 fps).
• The number of new frames added = total − conditioning frames. With ~13 tail frames conditioned and 81 total, you add ~68 new frames (~4 s) per pass.
• num_frames step is 4 in the node; keep total on the WAN 4n+1 grid (49 / 81 / 121 …). frame_rate for output is 16 fps (WAN 2.2 native).
• Resolution: 832×480 default (divisible by 16). ImageResizeKJv2 with crop/center and divisor 16 keeps the loaded clip on-grid.

---

Chaining multiple extensions (making a long video)

Pusa adds a bounded window (~4 s) per run. To go longer, feed the output back in:

0. Confirm the prior clip actually rendered — via the filesystem, not /history. VHS_VideoCombine writes the .mp4 but frequently does NOT register the output in ComfyUI's /history (the prompt shows done with no output and no error). Do NOT decide the render "silently dropped" from get_history / get_job_status alone — confirm the file with list_output_images (it now lists videos, each tagged kind: "video"): match the filename_prefix and check the mtime is fresh, then stage it.
1. Stage the output clip as the next run's input with stage_output_as_input (pass the rendered clip's { filename, subfolder?, type? }); use the returned input filename in VHS_LoadVideo. NEVER copy the output .mp4 into, or guess, a filesystem input/ path — ComfyUI's input/output dirs may be CUSTOM (--input-directory / --output-directory), so a guessed path makes VHS_LoadVideo fail to find/decode the file and wastes the run. The tool routes through the server API (/view → /upload/image), which resolves the real dirs correctly. (For a clip already on local disk, upload_video.)
2. Run the extension → decode → save (or keep the frames in-graph).
3. Take the tail of the new output (the last ~13 frames) as the next WanVideoEncode input.
4. Re-run with the same graph; the fresh tail becomes the new conditioning head.
5. Repeat. ImageConcatMulti / ImageBatchMulti (KJNodes, used in the example's preview) stitch the segments into one continuous clip.

Practical chaining tips:

• Always condition on the newest frames, not the original clip, or you'll "rewind."
• Drift compounds across hops (color/identity slowly wander). Keep noise_multipliers modest and re-state the subject in the prompt each hop.
• Overlap a few frames between segments and drop duplicates at concat to hide the seam.
• Keep the same seed discipline (fixed or deliberately varied) so motion cadence stays consistent.
• Each hop is an independent generation — clear_vram is not needed between hops, but decode/cache long chains to disk so you don't hold every segment in VRAM.

---

VRAM tiers

Same envelope as wan-t2v-video (dual A14B fp8 + UMT5) — Pusa adds only ~1.9 GB of LoRA. Use the wrapper's offload tooling.

• WanVideoModelLoader quant fp8_e4m3fn_scaled, base precision fp16_fast, offload_device, sageattn (the example's settings).
• Always clear_vram before switching to this from another model family.
• Encoder VAE tiling (WanVideoEncode) matters here because you're VAE-encoding real footage in addition to decoding output.

---

Gotchas

• Loading the example silently resets model/VAE/distill-LoRA dropdowns to the wrong first entry (subfolder paths don't resolve on a flat layout) — the #1 cause of a Pusa run that errors or generates wrong content. See "In practice: load → strip → re-point" and re-point ALL of them. Use strip_workflow to spot it.
• The prompt lives on WanVideoTextEncodeCached, not the CLIPTextEncode "decoy" branch (which isn't wired to the samplers).
• Interpolated source → seam speed jump — set VHS_LoadVideo.force_rate = 16, or condition on the pre-interpolation 16 fps master.
• sageattn / torch.compile errors — the example assumes SageAttention + triton. On a box without them, set attention_mode=sdpa and disconnect WanVideoTorchCompileSettings from both model loaders (TRAP 5).
• Saturation/color pop after the seam — re-grade the generated frames with a ColorMatchV2 (hm-mkl-hm) referencing the source clip, and use bf16 not fp16_fast (see "Seam color/saturation drift").
• Scheduler must be flowmatch_pusa. Leaving it on unipc/euler ignores the Pusa per-latent noise schedule → the conditioning latents don't behave as clean anchors and you get a hard cut / regeneration instead of a smooth continuation.
• merge_loras=false on fp8 (see CRITICAL above) — applies to the Pusa and distill LoRAs; default true silently kills the process.
• Match Pusa LoRA to expert: ...HIGH... → HIGH model, ...LOW... → LOW model. Crossing them degrades quality. Don't substitute the Wan 2.1 single-file rank512 LoRA into the 2.2 dual graph.
• Frame-count grid: keep num_frames on 4n+1 (49/81/121). Off-grid totals can error or pad oddly. num_frames UI step is 4.
• Motion drift / "frozen" continuation: tune noise_multipliers. Too low = stuck/looping; too high = subject/scene wanders. The 0→0.2 ramp is the safe middle.
• Color/exposure drift across chained hops is the most common long-video artifact. Mitigate: modest noise, restate the prompt, and optionally color-match each new segment to the previous before concat.
• Audio: WAN/Pusa generate silent video. The original clip's audio is not extended. Re-attach/curate audio at the end with VHS_VideoCombine (pass the source audio through) or in an editor — and note the new section has no native sound.
• ffmpeg required for the final mux (same as the other video skills): if VHS_VideoCombine errors ffmpeg ... could not be found, run <comfy-venv>/python -m pip install imageio-ffmpeg and reboot.
• Preview vs final VAE: taew2_1 (TinyVAE) is for fast preview decode; decode the final with the full WAN VAE for quality.

---

Minimal wiring (if you want fewer KJNodes)

The KJNodes (GetImageRangeFromBatch, GetLatentSizeAndCount, CreateScheduleFloatList, ImageResizeKJv2) are conveniences. The irreducible chain is:

load clip → (resize to 16-grid) → WanVideoEncode(vae, tail frames) → LATENT
WanVideoEmptyEmbeds(W,H,total) [extra_latents = that LATENT]  → embeds
embeds → WanVideoAddPusaNoise(noise_multipliers, noisy_steps) → embeds
WanVideoSampler(model+Pusa LoRA, embeds, scheduler=flowmatch_pusa, shift 5) → LATENT
WanVideoDecode(WAN VAE) → VHS_VideoCombine

You can hand a constant list to noise_multipliers instead of building a ramp; the ramp just smooths the seam. Two-pass HIGH→LOW is recommended (matches WAN 2.2's MoE) but a single LOW-model pass works for quick tests.

---

See also

• wan-t2v-video — the base WAN 2.2 T2V stack this builds on (model/encoder/ VAE loading, the merge_loras=false fp8 gotcha in full, block-swap/VRAM). Read it first.
• video-upscale — the natural next step: extend, then upscale. Generate / extend at 832×480, then run the result through the downscale → SeedVR2 (temporal restore+upscale) → RIFE → VHS encode pipeline for a clean, higher-res, higher-fps final. Do the extension first, upscale last (upscaling then extending wastes the restorer's work and risks re-drift).
• ltxv2-video — an alternative video family with its own extender variant; Pusa/WAN is the path when you want to continue an existing WAN-style clip.

Packs

No dedicated video-extend installer pack ships yet. Since Pusa reuses the installed WanVideoWrapper + KJNodes + VideoHelperSuite stack, a pack only needs to ensure those custom_nodes[] (kijai/ComfyUI-WanVideoWrapper, Kijai/ComfyUI-KJNodes, Kosinkadink/ComfyUI-VideoHelperSuite) plus the two Pusa V1 LoRAs in models[] (from Kijai/WanVideo_comfy/Pusa/). The big T2V A14B models are shared with wan-t2v-video — don't re-download. Install nodes ad-hoc with panel_install_node or apply a manifest with apply_manifest. Contribute a finished pack upstream (github.com/artokun/comfyui-mcp).