wanx-troopers
LongCat
2025.10.26
Resolutions supported?
ASPECT_RATIO_627 = {
'0.26': ([320, 1216], 1), '0.31': ([352, 1120], 1), '0.38': ([384, 1024], 1), '0.43': ([416, 960], 1),
'0.52': ([448, 864], 1), '0.58': ([480, 832], 1), '0.67': ([512, 768], 1), '0.74': ([544, 736], 1),
'0.86': ([576, 672], 1), '0.95': ([608, 640], 1), '1.05': ([640, 608], 1), '1.17': ([672, 576], 1),
'1.29': ([704, 544], 1), '1.35': ([736, 544], 1), '1.50': ([768, 512], 1), '1.67': ([800, 480], 1),
'1.73': ([832, 480], 1), '2.00': ([896, 448], 1), '2.31': ([960, 416], 1), '2.58': ([992, 384], 1),
'2.75': ([1056, 384], 1), '3.09': ([1088, 352], 1), '3.70': ([1184, 320], 1), '3.80': ([1216, 320], 1),
'3.90': ([1248, 320], 1), '4.00': ([1280, 320], 1)
}
I2V like with 5B or Pusa/Wan 5B, any number of start_frames as extra latents and the code replaces that part of the noise with the image, and sets corresponding timestep to 0 after expanding timesteps to be per frame
another WF screenshot
the exp args don’t do anything at cfg 1.0; left it there as the authors were using zero_star in their code when using cfg
full 50 steps to me
which sage version? use the 2.2.0 latest
Currently code in GH:kijai/ComfyUI-WanVideoWrapper:longcat
have to offload about half to run the bf16 on 4090; full offload and it’s under 10GB
their code does schedule like this when using the distll lora: 1000.0000, 994.7090, 988.7640, 982.0360, 974.3589, 965.5172, 955.2239, 943.0895, 923.0769, 904.1096, 880.3089, 849.5575, 808.2901, 750.0000, 661.4174, 510.6383 which is in the
longcat_distill_euler
Original News
Doordash released 14B parameters video AI model called LongCat, teaser.
MIT license, hurrah!
Promise of long and fast generations, 6 sec fragements stitched with 11 frame overlap; when continuation is used 11 frames is the only context possible, reference images are not taken.
In its own right as a base video generation model LongCat looks pretty nice too. Futher unlike Wan 2.2 it is one model not separate high/low.
example code does 832x480x93.
model structure is similar to Wan, just everything named differently, and has fused qkv; 48 blocks, smaller dim (4096); text encoding, VAE etc. are same; it’s not a Wan model
“First, the model generates a 480p, 15 fps video; second, this video is upscaled to 720p, 30f ps using trilinear interpolation and refined by a refinement expert”
refinement expert = lora, by the looks of it…
Cinescale style
And what’s amazing is that this Meituan is a food delivery platform.
An attempt to run in ComfyUI in wrapper, does not yet follow the prompt:
one model for T2V + I2V and continuations (similar to SVI) seems really exciting
everything would have to be trained, so not that easy to replace Wan: there’s no VACE, audio models, controlnets etc. etc.
it generates 6 second pieces in either t2v, i2v, or multi-frame i2v mode; I tried to generate a 12 second video in one go and it had the same repeating issue as wan
the multi frame i2v allows extension with motion continuity; those 11 [or 13 ?] frames are the only context it has for the next segment; if your character isnt perfectly in the shot at the boundary, rip
Apparenlty the authors provided an optional distill LoRa too but it is not functioning perfectly.
Long cat with torch compile turned off and sdpa instead of sageattn proved better in keeping the character consistent for me; Pixarom w/f
Community LoRa-s are missing
Refinement LoRa is for upscaling, distill LoRa is to speed up generation
LongCat prompt adherence is closer to Wan 2.1 than 2.2, it’s not great. Only thing it’s better at is video extension. And none of the Wan stuff like loras work on it so, not very useful model currently.