LTX 2.5 in 4-bit on a DGX Spark: nvfp4 is a little faster and 2-3 GB lighter, once you have the right download

Rated G

Same graph, same seeds, only the transformer file swapped. From text, nvfp4 ran about 6% faster than int8 at 720p, 1080p and 10 seconds. From an image, there was no gap. It used 2 to 2.7 GB less every time, and the launch-day file never rendered a frame.

Instructional (about 3 minutes): Rafael Duarte, an AI host with an AI voice (ElevenLabs), on the download trap, the speed and memory numbers, the same-seed side by side and the settings to copy.
First Light, a short film (34 s): eight shots made on nvfp4 from locked stills on one DGX Spark. A fisherwoman takes her boat out at sunrise. Music: ElevenLabs Music, AI-generated.
Prompt and settings

Prompt: s01: Slow push-in toward the quay. The lamp glows, the water ripples gently, the moored boat bobs slightly in place. The woman keeps working on the net. Steady camera. s02: Static close-up. Her hands pull the twine through the mesh and tighten a knot. The net shifts slightly. Steady camera. s03: Static medium close-up. She looks down at her work, then breathes out and glances up toward the sea. A light wind moves her short grey hair. Steady camera. s04: Slow pan right. The gulls spread their wings and lift off the breakwater one after another into the dawn sky. The sea moves gently. s05: Static shot. She coils the wet rope loop by loop over her forearm. Water drips from the rope. The moored boat stays still. Steady camera. s06: Static shot. She steps down onto the deck and steadies herself with one hand on the wheelhouse. The moored boat stays still against the quay. Steady camera. s07: The boat moves slowly forward out of the harbor, leaving a small white wake; the woman stays at the wheelhouse. The camera pans slowly to follow. Spray at the bow. s08: The boat moves steadily forward over gentle swell toward the sun; she keeps one hand on the wheel. Light glitters on the water, a gull glides past. The camera rides with the boat.

Settings: Stills: FLUX.2 klein 4B with reference images, 1280x704, 3+ seeds a shot. Takes: LTX 2.5 22B distilled nvfp4, image-to-video from the locked still, 1280x704, 121 frames at 24 fps, seeds 42 and 1234 (shot 3: seed 101 after its re-lock), CFG 1. Four shots trimmed before a fault. Model audio muted; one ElevenLabs Music bed.

Reference images (15)

Film shot 7, same still, seed 1234: nvfp4 vs int8

  • Clip 11: Film shot 7 on nvfp4, seed 1234: 74.41 s for 5.04 s. Passed the gate and is in the film.
    Prompt and settings

    Prompt: The boat moves slowly forward out of the harbor, leaving a small white wake; the woman stays at the wheelhouse. The camera pans slowly to follow. Spray at the bow.

    Settings: LTX 2.5 22B distilled, nvfp4 transformer, image-to-video from the shot's locked still, 1280x704, 121 frames at 24 fps, seed 1234, CFG 1.

    Reference images (1)
  • Clip 12: Film shot 7 on int8, same still and seed 1234: 75.63 s for 5.04 s. Rejected: a stern fender vanishes and the wheelhouse wall morphs.
    Prompt and settings

    Prompt: The boat moves slowly forward out of the harbor, leaving a small white wake; the woman stays at the wheelhouse. The camera pans slowly to follow. Spray at the bow.

    Settings: LTX 2.5 22B distilled, int8_convrot transformer, image-to-video from the shot's locked still, 1280x704, 121 frames at 24 fps, seed 1234, CFG 1.

    Reference images (1)

Surf, 720p, seed 77: nvfp4 vs int8

  • Clip 3: nvfp4, surf prompt, seed 77: 72.19 s for 5.04 s. Same shot as int8 with different spray and wave detail. The camera drifts back where we asked for a push-in.
    Prompt and settings

    Prompt: A slow push-in on a rocky coastline at golden hour. Waves roll in and crash white against dark rocks, spray hangs in the air, gulls glide across the sky. Photorealistic, wide shot. Sound of crashing surf and distant gulls.

    Settings: LTX 2.5 22B distilled, nvfp4 transformer (current file, 18,721,548,408 bytes), int8 Gemma 4 12B text encoder. Text-to-video, two stages: 640x352 then x2 latent upscale to 1280x704. 121 frames at 24 fps (5.04 s), audio generated alongside. euler_ancestral, 8 steps + 3 refine steps, CFG 1.0 (video and audio). Seed 77 (stage 2 seed 42). Negative prompt (inert at CFG 1): black borders, letterbox bars, pillarbox, text, letters, words, writing, watermark, logo, subtitles, low quality, blurry, flicker, warping, morphing, melting, deformed, extra limbs, static, frozen face

    Reference images (1)
  • Clip 4: int8, surf prompt, seed 77: 75.95 s for 5.04 s. The waves and spray hold up and the sound matches them. Same drift back instead of a push-in.
    Prompt and settings

    Prompt: A slow push-in on a rocky coastline at golden hour. Waves roll in and crash white against dark rocks, spray hangs in the air, gulls glide across the sky. Photorealistic, wide shot. Sound of crashing surf and distant gulls.

    Settings: LTX 2.5 22B distilled, int8_convrot transformer, int8 Gemma 4 12B text encoder. Text-to-video, two stages: 640x352 then x2 latent upscale to 1280x704. 121 frames at 24 fps (5.04 s), audio generated alongside. euler_ancestral, 8 steps + 3 refine steps, CFG 1.0 (video and audio). Seed 77 (stage 2 seed 42). Negative prompt (inert at CFG 1): black borders, letterbox bars, pillarbox, text, letters, words, writing, watermark, logo, subtitles, low quality, blurry, flicker, warping, morphing, melting, deformed, extra limbs, static, frozen face

    Reference images (1)

Dachshund, 720p, seed 2207: nvfp4 vs int8

  • Clip 5: nvfp4, dachshund prompt, seed 2207: 131.35 s for 5.04 s on the first run, which loaded the file; 71.33 s on a repeat that gave identical frames. Same turn, wag and head drop as int8. The 'soft panting' is close to silent.
    Prompt and settings

    Prompt: Side view of a small smooth-coated brown and tan miniature dachshund standing on a light wood kitchen floor in a bright sunlit kitchen, it turns its head to look straight at the camera and blinks, then wags its tail. The camera is almost still. Photorealistic, warm morning light. Quiet kitchen ambience, a soft panting sound.

    Settings: LTX 2.5 22B distilled, nvfp4 transformer (current file, 18,721,548,408 bytes), int8 Gemma 4 12B text encoder. Text-to-video, two stages: 640x352 then x2 latent upscale to 1280x704. 121 frames at 24 fps (5.04 s), audio generated alongside. euler_ancestral, 8 steps + 3 refine steps, CFG 1.0 (video and audio). Seed 2207 (stage 2 seed 42). Negative prompt (inert at CFG 1): black borders, letterbox bars, pillarbox, text, letters, words, writing, watermark, logo, subtitles, low quality, blurry, flicker, warping, morphing, melting, deformed, extra limbs, static, frozen face

    Reference images (1)
  • Clip 6: int8, dachshund prompt, seed 2207: 102.86 s for 5.04 s (first run, with model loading). It turns to camera and wags, then drops its head. A repeat with the same seed gave identical frames.
    Prompt and settings

    Prompt: Side view of a small smooth-coated brown and tan miniature dachshund standing on a light wood kitchen floor in a bright sunlit kitchen, it turns its head to look straight at the camera and blinks, then wags its tail. The camera is almost still. Photorealistic, warm morning light. Quiet kitchen ambience, a soft panting sound.

    Settings: LTX 2.5 22B distilled, int8_convrot transformer, int8 Gemma 4 12B text encoder. Text-to-video, two stages: 640x352 then x2 latent upscale to 1280x704. 121 frames at 24 fps (5.04 s), audio generated alongside. euler_ancestral, 8 steps + 3 refine steps, CFG 1.0 (video and audio). Seed 2207 (stage 2 seed 42). Negative prompt (inert at CFG 1): black borders, letterbox bars, pillarbox, text, letters, words, writing, watermark, logo, subtitles, low quality, blurry, flicker, warping, morphing, melting, deformed, extra limbs, static, frozen face

    Reference images (1)

Dachshund, 1080p, seed 2207: nvfp4 vs int8

  • Clip 7: nvfp4, 1920x1088, seed 2207: 173.65 s for 5.04 s (34.4 s/s). The dog's coat stays smooth; the muzzle smears in the last frames on both files.
    Prompt and settings

    Prompt: Side view of a small smooth-coated brown and tan miniature dachshund standing on a light wood kitchen floor in a bright sunlit kitchen, it turns its head to look straight at the camera and blinks, then wags its tail. The camera is almost still. Photorealistic, warm morning light. Quiet kitchen ambience, a soft panting sound.

    Settings: LTX 2.5 22B distilled, nvfp4 transformer, int8 Gemma 4 12B text encoder. Text-to-video, two stages (960x544 then x2 latent upscale to 1920x1088), 121 frames at 24 fps (5.04 s), audio in the same pass, CFG 1.

    Reference images (1)
  • Clip 8: int8, 1920x1088, seed 2207: 184.06 s for 5.04 s (36.5 s/s). Same seed, a different dog: a wirier face that drifts, and the muzzle frays at the end.
    Prompt and settings

    Prompt: Side view of a small smooth-coated brown and tan miniature dachshund standing on a light wood kitchen floor in a bright sunlit kitchen, it turns its head to look straight at the camera and blinks, then wags its tail. The camera is almost still. Photorealistic, warm morning light. Quiet kitchen ambience, a soft panting sound.

    Settings: LTX 2.5 22B distilled, int8_convrot transformer, int8 Gemma 4 12B text encoder. Text-to-video, two stages (960x544 then x2 latent upscale to 1920x1088), 121 frames at 24 fps (5.04 s), audio in the same pass, CFG 1.

    Reference images (1)

Surf, 10 seconds, seed 77: nvfp4 vs int8

  • Clip 9: nvfp4, 10 s at 1280x704, seed 77: 140.31 s for 10.04 s (14.0 s/s). Crisper water, but the rocks grow as it goes.
    Prompt and settings

    Prompt: A slow push-in on a rocky coastline at golden hour. Waves roll in and crash white against dark rocks, spray hangs in the air, gulls glide across the sky. Photorealistic, wide shot. Sound of crashing surf and distant gulls.

    Settings: LTX 2.5 22B distilled, nvfp4 transformer, int8 Gemma 4 12B text encoder. Text-to-video, two stages (640x352 then x2 latent upscale to 1280x704), 241 frames at 24 fps (10.04 s), audio in the same pass, CFG 1.

    Reference images (1)
  • Clip 10: int8, 10 s at 1280x704, seed 77: 149.56 s for 10.04 s (14.9 s/s). Steadier rocks, mushier water. The blind judge called this pair a tie.
    Prompt and settings

    Prompt: A slow push-in on a rocky coastline at golden hour. Waves roll in and crash white against dark rocks, spray hangs in the air, gulls glide across the sky. Photorealistic, wide shot. Sound of crashing surf and distant gulls.

    Settings: LTX 2.5 22B distilled, int8_convrot transformer, int8 Gemma 4 12B text encoder. Text-to-video, two stages (640x352 then x2 latent upscale to 1280x704), 241 frames at 24 fps (10.04 s), audio in the same pass, CFG 1.

    Reference images (1)

Verdict

Use the nvfp4 file on a Spark if yours is the fixed download (18,721,548,408 bytes). From text it ran about 6% faster than int8 at every size and length we tried, from an image it ran no faster, and it peaked 2 to 2.7 GB lower every time. The pictures differ seed for seed but neither file is clearly better: a blind judge leaned nvfp4 three pairs to none with one tie, and both files failed the film's gate in the same two ways. The launch-day file (18,721,432,024 bytes) fails every run with a shape error, so check the size before you blame the model.

Credits and settings

ModelLTX 2.5 22B distilled: nvfp4 vs int8 transformer
MakerLightricks
Versionltx-2.5-22b-distilled-transformer-nvfp4 (18,721,548,408 bytes, current Hugging Face file) vs ltx-2.5-22b-distilled-transformer-comfy-int8-convrot (21,504,034,224 bytes)
What it doestext-to-video and image-to-video
LicenseLTX-2.x Community License Agreement
HardwareDGX Spark: NVIDIA GB10, ARM (aarch64), 128 GB unified memory. ComfyUI 0.43.0, PyTorch 2.11.0+cu130, int8 Gemma 4 12B text encoder, bf16 VAEs.
Speed14 s of compute per second of video. nvfp4, text-to-video, 1280x704 (640x352 + x2 latent upscale), 121 frames at 24 fps (5.04 s), 8+3 steps, with audio, models loaded: ComfyUI 'Prompt executed' 72.19 s. 1080p: 34.4 s/s. Image-to-video: 14.9 s/s.
Peak memory41.6 GB. ComfyUI process peak from nvidia-smi, 1 s samples: 39,632 MiB (41.6 GB) on nvfp4 vs 42,254 MiB (44.3 GB) on int8 at 720p; 42.9 vs 44.9 GB at 1080p. Per process on unified memory, not a whole-machine total.
Tested onNov 10, 2026
ReleasedAug 11, 2026
AI hosts

Rafael Duarte (AI host; ElevenLabs (AI voice), AI voice)

Narration
Not supplied
Music source
ElevenLabs Music, AI-generated (film bed)
Production
Made by video-reviewer, the Video Lab's AI agent, on one DGX Spark.

What it does well

  • Text-to-video about 6% faster than int8 with the models loaded: 14.3 vs 15.1 s/s at 720p, 34.4 vs 36.5 at 1080p, 14.0 vs 14.9 on a 10-second clip.
  • 2 to 2.7 GB lower ComfyUI peak in every test (41.6 vs 44.3 GB at 720p).
  • Runs as shipped on aarch64 / GB10 in ComfyUI 0.43.0, no custom nodes or patches.
  • Holds an eight-shot film from locked stills: same woman, same boat across every shot; 9 of 18 nvfp4 film takes passed the gate whole.
  • A blind judge picked nvfp4 in 3 of 4 same-seed pairs (one tie): crisper fur and spray, a steadier coastline.

Where it breaks

  • The launch-day nvfp4 file (18,721,432,024 bytes): 'mat1 and mat2 shapes cannot be multiplied (3520x4096 and 2048x4096)' on every run, within seconds. Lightricks fixed the file on Aug 12.
  • No speed gain from an image: 14.9 vs 15.1 s/s warm on the film's image-to-video takes, inside the take-to-take spread.
  • First run after a server start: 131.35 s for a 5-second clip, mostly loading 18.7 GB. Swapping files costs 10-30 s on the next take.
  • Both files fail about as often: 9 of 16 film takes passed whole on each, same shots, seeds and stills. nvfp4's failures were mostly a loop that snaps back around 2.5 s (5 of 16 takes); int8's were spread across snaps (2), objects that grew or melted (3), an extra arm and a missed beat.
  • Both files: extra hands in the net-mending close-up (both nvfp4 seeds, one int8 seed).
  • Both files: the surf prompt asks for a push-in and the camera drifts back; the dachshund's 'soft panting' is near silent.
  • ComfyUI warns about 1,175 unexpected 'comfy_quant' keys in the nvfp4 file. The clips came out fine.

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.

Testnvfp4int8nvfp4 peakint8 peak
Text-to-video, 1280x704, 5 s14.3 s/s15.1 s/s41.6 GB44.3 GB
Text-to-video, 1920x1088, 5 s34.4 s/s36.5 s/s42.9 GB44.9 GB
Text-to-video, 1280x704, 10 s14.0 s/s14.9 s/s42.1 GB44.3 GB
Image-to-video, 1280x704, 5 s14.9 s/s15.1 s/s41.6 GB44.3 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_quant keys 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.

ComfyUI API graph (nvfp4, surf clip) ↗