The film at the top, First Light, is eight shots made on nvfp4 from locked stills. Two of its shots were also rendered on int8 from the same stills and seeds, so you can see what changes when only the file does.
How we tested
One ComfyUI graph, fixed prompts and seeds, and only the transformer file swapped between runs: the Comfy int8 "convrot" file (21.5 GB) or the nvfp4 file (18.7 GB). The GB10 is a Blackwell chip and nvfp4 is the 4-bit format Blackwell has hardware for, which is why it's worth asking.
| Test | nvfp4 | int8 | nvfp4 peak | int8 peak |
|---|---|---|---|---|
| Text-to-video, 1280x704, 5 s | 14.3 s/s | 15.1 s/s | 41.6 GB | 44.3 GB |
| Text-to-video, 1920x1088, 5 s | 34.4 s/s | 36.5 s/s | 42.9 GB | 44.9 GB |
| Text-to-video, 1280x704, 10 s | 14.0 s/s | 14.9 s/s | 42.1 GB | 44.3 GB |
| Image-to-video, 1280x704, 5 s | 14.9 s/s | 15.1 s/s | 41.6 GB | 44.3 GB |
s/s is seconds of compute per second of finished video, with the models already loaded. Peaks are the ComfyUI process.
What that means in practice:
- From text, nvfp4 is about 6% faster, and the gap doesn't grow. It's the same at 1080p and at 10 seconds as it is at 720p and 5 seconds. In real time that's about 4 seconds saved on a 5-second 720p clip and 10 seconds on a 1080p one: twenty 1080p clips, three and a half minutes.
- From an image, there's no gap we could measure. Twelve warm nvfp4 takes averaged 75.3 s, three int8 takes 75.9 s, and the spread between takes of the same file was about 9 s. Same shot, same seed: 75.05 against 74.24, 79.01 against 77.83, 74.41 against 75.63.
- Memory is the steady win: 2 to 2.7 GB lower in every test. That matters if you keep a second model loaded next to LTX.
- 1080p costs about 2.4 times what 720p does per second of video. A 10-second clip costs the same per second as a 5-second one.
Why so small? On the Spark, video generation is limited by compute, not memory bandwidth, and right now int8 runs on a slow path: an open comfy-kitchen issue (#280) measures int8 on the GB10 at about 37% of what the chip can do. If that gets fixed, int8 could pass nvfp4. Run the 720p test again after you update ComfyUI.
Does it look different?
Same seed, different file, different clip. The side-by-sides below show it.
- Blind pick. Shown four same-seed pairs without the labels, a judge picked nvfp4 three times (crisper fur and spray, a steadier coastline, a cleaner face at 1080p) and called the 10-second surf clip a tie. Call it a lean toward nvfp4.
- The film's gate. Every film take went through the same check: same woman, same boat, nothing appearing or melting. 9 of 18 nvfp4 takes passed whole. To compare the files fairly, we then ran int8 on all eight shots with the same two seeds and the same stills. 9 of 16 passed on each file, so the rate is a tie. The failures weren't the same mix. Five of nvfp4's seven were a loop that snaps back around 2.5 seconds, and the other two were extra hands in the net-mending close-up. int8's seven were spread out: two of the same snaps, three objects that grew or melted (fenders, a wheelhouse wall, a rope), an extra arm, and one take where she never stepped aboard. Sixteen takes a file is enough to call the rate even, not enough to call either file cleaner. Every one of these failures shows up on int8 too, so none of them is the 4-bit file's fault.
The download trap
The nvfp4 file we downloaded on launch day (Aug 11) never rendered a frame. Every run died within seconds with mat1 and mat2 shapes cannot be multiplied (3520x4096 and 2048x4096), because ComfyUI didn't recognize the file as quantized and read the packed 4-bit weights as an ordinary matrix. Lightricks replaced the file on Aug 12.
If you grabbed nvfp4 early and it fails like that, it isn't your Spark. The file is ltx-2.5-22b-distilled-transformer-nvfp4.safetensors on Lightricks/LTX-2.5. Check yours with stat -c %s ltx-2.5-22b-distilled-transformer-nvfp4.safetensors in ComfyUI's model folder:
- Fixed file: 18,721,548,408 bytes. Works.
- Launch-day file: 18,721,432,024 bytes. Fails on every run.
How we made First Light
A fisherwoman mends her net on the quay before sunrise, coils her rope and takes her red boat out as the sun comes up. We picked it for the details 4-bit could smear: net mesh, wet rope, spray, gull wings, low sun through mist.
We made a reference picture of the woman, the boat and the harbor first, then every shot's still from those references with FLUX.2 klein, three seeds a shot. 18 of 28 stills passed the check, and one of those shouldn't have. Shot 8 failed twice the same way (the boat at sea with nobody at the wheel), so we reframed it from the stern with her at the wheel instead of rolling more seeds. Each locked still went to LTX 2.5 on nvfp4, two seeds a shot.
The one that slipped through was shot 3, the close-up of her face. All three seeds gave her an older man's face instead of the woman in the reference, and the still passed because the jacket and the hair matched. Our QA gate caught it after the film was cut. The prompt led with "weathered" and "tanned lined face", and klein read that as a man. We rewrote it to describe her face first (a woman of about 60, a rounded face, small earrings) and all three new seeds came back as her. Of two new takes, one passed and one failed because she turned her back to the camera. If a character's face drifts in your stills, fix the words before you roll more seeds.
Four shots in the film are trims, cut just before the fault: a second hand at 1.2 s in the net close-up, the 2.5-second snap in two others, and a small twitch of her head at 4.6 s in the face close-up.
Setup and gotchas
- ComfyUI 0.43.0, PyTorch 2.11.0+cu130 on aarch64. Both files load with no custom nodes or patches.
- The first nvfp4 run after a server start took 131 s for a 5-second clip, mostly reading 18.7 GB from disk. Count on that once.
- Swapping between the two files costs time too: the first take after a swap ran 84 to 105 s instead of about 75.
- ComfyUI logs a warning about 1,175
comfy_quantkeys in the nvfp4 file. The clips came out fine. - Load the full VAE. One Spark write-up traced its soft nvfp4 clips to a tiny preview VAE, and found that an nvfp4 checkpoint with the VAE built in can hang on the GB10.
Settings
- LTX 2.5 22B distilled, transformer
ltx-2.5-22b-distilled-transformer-nvfp4(or the int8 convrot file; nothing else changes). - Text encoder
gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot; video and audio VAEs bf16; latent spatial upscaler x2 bf16. - Two stages: 640x352 for 8 steps, x2 latent upscale, then 3 refine steps at 1280x704 (1920x1088 for the 1080p test).
- 121 frames at 24 fps (241 for the 10-second test), audio in the same pass.
- CFG 1.0. The distilled model ignores the negative prompt at CFG 1, so put what you want in the prompt.
- The workflow file below is the exact graph for the nvfp4 surf clip.
Speed and memory
The table's times cover the whole graph, text to finished file, with the models loaded. Memory is what the ComfyUI process held at its peak, not the whole machine, and the Spark never had less than 35.7 GB free, so there's room to run something next to it.