DEV Community

AI-KSK
AI-KSK

Posted on

LongCat Avatar All-in-One: One Graph for Audio Drive, Loop Continuation and Multi-Image Switching

#comfyui #ai #tutorial

Instead of switching graphs between a single-shot digital human, a loop and a multi-image music-video style cut, this episode packs all three into one switch-driven graph - 210 nodes for the single-image loop version and 197 for the multi-shot version. The switches are the interface: prompt mode, reference mode, loop on/off, loop count, cut mode.

Evidence note: two workflow JSONs ship with this episode; the WanVideo sampler, scheduler, LoRA and VAE values below are read from them. Defaults such as 93 frames and 13 overlap frames are stated by the author and also present in the graphs.

When to use it

  • You want a talking presenter that loops indefinitely with one image.
  • You want hard cuts between different images (MV style) instead of a soft continuation.
  • You want 720p and are willing to accept a fixed segment size.

What is different in this workflow

  • Switch-driven instead of graph-switching: prompt mode, reference mode, loop switch, loop count and cut mode are all numeric toggles.
  • Two latent-continuation modes: hard cut (0) forces the previous latent to be the encoded reference image; soft continuation (1) uses the last segment's tail latent, which lowers the reference latent's influence.
  • Frame budget: 93 frames for the first segment, +80 frames per loop.
  • Auto loop count: loop mode 0 derives the count from audio length; mode 1 takes a manual number.
  • Prompt/image rotation: 0 fixes, 1 rotates through up to 10 slots, so one graph covers both a fixed presenter and a cutting MV.

Step by step

  1. Load audio and trim the start time (the author only wants from second 19 in his demo).
  2. Set width/height: 720x1280 or 1280x720 are the official picks; the author says 720p is where LongCat earns its keep.
  3. Keep 93 frames per segment and 13 overlap frames for a single-image loop; set overlap to 1 for multi-image cuts.
  4. Keep CFG at 1 - this is already an accelerated graph.
  5. Prompt mode 0 = fixed, 1 = rotate; reference mode 0 = fixed first image, 1 = rotate by index through 1-10.
  6. Loop switch 0 = one segment only (93 frames); 1 = loop, +80 frames per loop. Loop count 0 = derive from audio, 1 = manual.
  7. Cut mode: 0 soft continuation, 1 hard cut. For the MV-style hard cuts use 1 with overlap 1.
  8. Reference images: up to 10 can be listed; one is enough for a fixed presenter. Captions under each image are reverse-written by GPT-5; set that mode to 0 to keep only the first caption, 1 to pass them in order.
  9. Preset for "one image, infinite loop": prompt mode 0, reference mode 0, loop switch 1, loop count 0 (or 1 for manual), cut mode 0.

Parameter basis

  • Graphs: LongCat Avatar 单图循环数字人.json (210 nodes) and LongCat Avatar 多图切镜数字人.json (197 nodes).
  • Model: WanVideoModelLoader -> LongCat-Avatar-15_bf16.safetensors, dtype bf16, block swap disabled, offload to device, attention sageattn.
  • LoRA: WanVideoLoraSelect -> LongCat-Avatar-15_dmd_distill_lora_rank128_bf16.safetensors at 0.9.
  • VAE: WanVideoVAELoader -> Wan2_1_VAE_bf16.safetensors.
  • Text encoder: WanVideoTextEncodeCached -> umt5-xxl-enc-bf16.safetensors.
  • Scheduler: WanVideoSchedulerv2 -> longcat_distill_euler, 8 steps, shift 12, start 0, end -1.
  • Sampler: WanVideoSampler at 1 step, CFG 1, fixed seed (two passes, seeds 1 and 2).
  • Continuation: WanVideoLongCatAvatarExtendEmbeds with 93 frames / 13 overlap / index 0 or 93 / pad_with_start / 10 / 3.
  • Decode: WanVideoDecode with 272x272x144 chunking.
  • Frame rate: 25 fps (author states 25 in the video).

Limits

  • 720p is the practical ceiling; the author says anything lower loses LongCat's advantage.
  • Hard-cut mode lowers how much the reference latent matters, so identity is carried by the loop overlap rather than by the image.
  • 10 reference slots is not a loop limit: with a fixed image and fixed prompts you can loop far more times.
  • Freeform prompt variation means deleting the reverse-prompt node and typing by hand - the author is explicit that it is not as plug-and-play at that point.
  • The graph is big (197-210 nodes); most of the nodes are annotation notes.

Episode resources

Top comments (0)

⌂ 换网址