Video Lab · text-to-video

LTX 2.5 22B distilled on the Spark: solid 1080p shots at 36 s per second of video, as long as the camera barely moves

Ten text-to-video clips from an Oct 31 product-shot batch: what the int8 distilled model holds together, where it ignores the prompt, and why the negative prompt did nothing.

Tested on Jon's DGX Spark · Oct 31, 2026

The best clip of the batch. A dachshund in profile turns its head to camera, holds, and blinks at the end, exactly as asked. Fur, floor reflections and the blurred kitchen hold steady for all five seconds. Wall time 183 s.

Prompt and settings

Prompt: Side view of a small smooth-coated brown and tan miniature dachshund standing still on a light wood kitchen floor in a bright sunlit kitchen, long body in profile filling the middle of the frame, it turns its head to look straight at the camera with a proud serious expression and blinks. Almost motionless. The camera is almost still. Full-frame image filling the entire frame edge to edge, photorealistic, warm morning light.

Settings: text-to-video, 1920x1088, 121 frames at 24 fps (5.04 s), seed 2207, 8 + 3 steps, CFG 1.0

Same dog, seated, with a chin lift and head tilt. The motion is there but the body reads as a glossy figurine more than a real dog, and the sage cabinets behind it are the only part of the prompt's kitchen that showed up. Wall time 186 s.

Prompt and settings

Prompt: A small smooth-coated red-brown miniature dachshund sits very still on a clean light oak kitchen floor in a bright sunny modern kitchen with sage-green cabinets, its long body turned sideways to the camera so its whole side is visible, head turned toward the camera. It proudly lifts its chin and slowly tilts its head, ears flopping slightly. Otherwise almost motionless. The camera is almost still. Full-frame image filling the entire frame edge to edge, photorealistic, warm natural light, shallow depth of field.

Settings: text-to-video, 1920x1088, 121 frames at 24 fps (5.04 s), seed 1105, 8 + 3 steps, CFG 1.0

The one real camera move that worked: a slow sideways glide along pantry shelves, with parallax between the bins and the posts. Every label is a white rectangle of pseudo-text. Wall time 144 s.

Prompt and settings

Prompt: A beautifully organized pantry with light oak shelves holding rows of matching clear bins and glass jars of pasta, rice, oats, lentils and snacks, each with a small neat white rectangular label. Warm morning sunlight streams in. The camera glides very slowly sideways along the shelves. Full-frame image filling the entire frame edge to edge, photorealistic home lifestyle commercial, cream and sage palette.

Settings: text-to-video, 1920x1088, 97 frames at 24 fps (4.04 s), seed 909, 8 + 3 steps, CFG 1.0

Tungsten-lit hand writing on a freezer bag. Hand, marker and bag stay solid for four seconds; the writing itself is scribble, which is what the prompt asked for. Wall time 147 s.

Prompt and settings

Prompt: Close-up at a wooden kitchen table in the evening: a hand slowly writes with a black permanent marker on a clear plastic zip freezer bag of raw chicken breasts, scribbling a messy handwritten label. More freezer bags of ground beef are stacked nearby, a marker cap on the table. Warm tungsten lamp light, slightly tired mood, shallow depth of field. The camera is almost still. Full-frame image filling the entire frame edge to edge, photorealistic cinematic.

Settings: text-to-video, 1920x1088, 97 frames at 24 fps (4.04 s), seed 606, 8 + 3 steps, CFG 1.0

Overhead table of freezer bags. The marker rolls, the bags sit still, the lamp light is right. The printed labels on the bags are gibberish glyphs. Wall time 114 s for 3 s of video.

Prompt and settings

Prompt: Overhead view of a cluttered wooden kitchen table covered with a dozen clear zip freezer bags of portioned raw chicken and ground beef, smudged messy marker scribbles on them, an uncapped black marker rolling slowly across the table. Warm evening lamp light. The camera is almost still, very slow drift. Full-frame image filling the entire frame edge to edge, photorealistic cinematic.

Settings: text-to-video, 1920x1088, 73 frames at 24 fps (3.04 s), seed 707, 8 + 3 steps, CFG 1.0

Asked for a drawer sliding open toward the camera. Got a near-static shot of neatly packed bags with barcode-like fake labels; there is a slight drift, no slide. The cold palette and frosted film look are convincing. Wall time 147 s.

Prompt and settings

Prompt: A freezer drawer slides open toward the camera revealing perfectly organized rows of flat clear freezer bags of portioned chicken, ground beef and soup, each with a small neat white rectangular sticker label. Soft cold light with a little frosty mist drifting out, clean and satisfying. Full-frame image filling the entire frame edge to edge, photorealistic product commercial.

Settings: text-to-video, 1920x1088, 97 frames at 24 fps (4.04 s), seed 808, 8 + 3 steps, CFG 1.0

Asked for a door swinging open. Got a wall of labeled jars with a faint shadow passing across the top of the frame, and a sliver of produce drifting in at the right edge. The big motion in the prompt did not happen. Wall time 144 s.

Prompt and settings

Prompt: A tall modern refrigerator door swings open toward the camera revealing a perfectly organized, sparkling clean, brightly lit fridge: rows of matching clear glass containers, jars and produce bins, each with a small neat white rectangular label, everything lined up flawlessly. Soft cool fridge light spills out. Full-frame image filling the entire frame edge to edge, photorealistic product commercial.

Settings: text-to-video, 1920x1088, 97 frames at 24 fps (4.04 s), seed 3404, 8 + 3 steps, CFG 1.0

"Washed-out low-saturation infomercial" came out as a muted but still colour clip. The marker skips and scribbles as asked, the hand stays intact. Wall time 144 s.

Prompt and settings

Prompt: Washed-out low-saturation 1990s late-night TV infomercial footage. Extreme close-up of a woman's hand pressing a black permanent marker onto a wet, frosty plastic freezer bag full of frozen chicken; the marker skips and leaves only faint broken streaks, she scribbles harder and harder in frustration, the bag slides around on the counter. Flat bright studio lighting. Full-frame image filling the entire frame edge to edge.

Settings: text-to-video, 1920x1088, 97 frames at 24 fps (4.04 s), seed 3101, 8 + 3 steps, CFG 1.0

Same "low-saturation" wording, and this time the model went full black and white. The spilled peas bounce and scatter with believable physics, which is the best motion in the 4 s clips. Hands blur in the fast frames. Wall time 144 s.

Prompt and settings

Prompt: Washed-out low-saturation 1990s late-night TV infomercial footage of a chaotic kitchen counter buried in uncapped markers, rolls of masking tape, torn tape strips and crumpled ziplock bags; a ziplock bag tips over and spills frozen peas and rice everywhere across the counter and onto the floor while a woman's hands flail trying to catch it. Exaggerated, comedic. Flat bright lighting. Full-frame image filling the entire frame edge to edge.

Settings: text-to-video, 1920x1088, 97 frames at 24 fps (4.04 s), seed 3202, 8 + 3 steps, CFG 1.0

Black and white again. Greasy hands fumble a label maker and drop it; the fingers smear into a blur during the drop and the label maker's keys are a mush of shapes. Wall time 144 s.

Prompt and settings

Prompt: Washed-out low-saturation 1990s late-night TV infomercial footage. Close-up of a woman's hands covered in shiny cooking grease and flour clumsily jabbing the tiny keys of a small handheld label maker; it slips out of her slippery fingers and clatters onto the messy counter. Comedic, exaggerated frustration. Flat bright lighting. Full-frame image filling the entire frame edge to edge.

Settings: text-to-video, 1920x1088, 97 frames at 24 fps (4.04 s), seed 3303, 8 + 3 steps, CFG 1.0

Verdict

Run it on a Spark for still-camera product and lifestyle shots at 1920x1088: about three minutes per five-second clip, the subject stays solid, and a single small motion (a head turn, a hand writing, something spilling) comes out clean. Don't rely on it for the big move in your prompt, for any text in frame, or for a style word like "low-saturation", which it took as colour once and black-and-white twice. The distilled checkpoint runs at CFG 1.0, so your negative prompt is decoration.

Model	LTX 2.5 22B distilled
Maker	Lightricks
Version	22B distilled transformer, Comfy int8 convrot (ltx-2.5-22b-distilled-transformer-comfy-int8-convrot), with the int8 Gemma 4 12B text encoder
What it does	text-to-video
License	LTX-2.x Community License Agreement
Hardware	DGX Spark: NVIDIA GB10, ARM (aarch64), 128 GB unified memory. ComfyUI 0.43.0, PyTorch 2.11.0+cu130, int8 transformer and text encoder, bf16 VAEs.
Speed	36 s of compute per second of video. 1920x1088 text-to-video, 24 fps, two stages (8 steps at 960x544, 2x latent upscale, 3 steps at full size), wall time from submit to done with weights already loaded. Image-to-video with the same graph: about 42.
Peak memory	38.5 GB. nvidia-smi per-process figure for the ComfyUI holding the weights after the run (matches the 38.1 GB the server log staged). Peak during sampling not measured.
Tested on	Oct 31, 2026
Released	Aug 11, 2026

What it does well

Subjects hold their shape for the full clip: the dachshund, the writing hand and the freezer bags never melt or morph in these ten clips
Small motions land: the head turn and blink in dachshund-b, the marker in sharpie-before, the pea spill in counter-chaos
Lighting and palette from the prompt come through (tungsten evening, cold freezer light, warm morning kitchen)
Slow lateral camera glides work, with real parallax (pantry)
Predictable cost: 36 s of compute per second of 1920x1088 video, within a few seconds across all ten t2v clips

Where it breaks

Can't write: every label, sticker and barcode is pseudo-text, in all ten clips
Drops the big action: 'drawer slides open' (freezer) and 'door swings open' (fridge-open) both came out as near-static shots
Style words are a coin flip: 'washed-out low-saturation' gave muted colour in sharpie-fail and full black-and-white in counter-chaos and greasy-labelmaker
Fast hands smear: fingers blur during the drop in greasy-labelmaker and the flail in counter-chaos
The dog in dachshund-a reads as a glossy figurine rather than fur
Negative prompt has no effect at CFG 1.0, so you can't negative-prompt your way out of any of the above
Output files are silent unless you wire the audio latent into CreateVideo yourself

How we tested

This write-up reports a run we did on this machine on Oct 31, 2026, not a fresh test. The GPU was busy today (85 GB of the 128 GB held by other resident models), so nothing new was rendered. Everything below comes from that run's batch logs, the ComfyUI server log, the workflow JSON files, and the clips themselves.

The job was a set of product-commercial shots: 15 clips, all 1920x1088 at 24 fps, 3 to 5 seconds each, submitted one at a time to ComfyUI through its API by a small Python script. Five of them were image-to-video from branded stills for a product that isn't public yet, so those clips stay off this page. The ten text-to-video clips are above, untouched apart from a remux so they stream (the originals are the H.264 files ComfyUI wrote, 1.3 to 3.7 MB each).

What the ten clips show, in short: the model is dependable for a still or near-still camera with one small subject motion (the dog turning its head, a hand writing, peas spilling). It's unreliable for the big move in a prompt (a drawer sliding open, a fridge door swinging) and it can't write: every label in every clip is pseudo-text.

Setup and gotchas

ComfyUI 0.43.0 or newer. The LTX 2.5 video VAE and the Gemma 4 text encoder won't load on 0.27; that was the first thing that broke on this machine. We run 0.43.0 (commit 426f5ee, 2026-10-26) with PyTorch 2.11.0+cu130 on driver 580.173.02.
Files, all from the Comfy-Org repack: transformer ltx-2.5-22b-distilled-transformer-comfy-int8-convrot (21.5 GB), text encoder gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot (15.4 GB), video VAE bf16 (1.5 GB), audio VAE bf16 (0.4 GB), latent spatial upscaler x2 bf16 (1.0 GB). About 39 GB on disk.
The int8 "convrot" transformer loads through ComfyUI's mixed-precision ops. The log shows "Native ops: w6a8_int8, convrot_w4a4, ..." and "model weight dtype torch.bfloat16". No custom nodes, no patches; it ran on aarch64 as shipped.
CFG 1.0 makes the negative prompt inert. The distilled checkpoint is a fixed-schedule, CFG=1 model. At CFG 1.0 the guider's output equals the positive-only prediction, so the long negative prompt in this workflow ("no text, no letters, no borders...") did nothing. The gibberish labels in the pantry, freezer and fridge clips are the proof.
No audio in the output files. The graph samples an audio latent alongside the video (that's what LTX 2.5 does) but the CreateVideo node was never given the decoded audio, so every mp4 is silent. If you want the model's audio, decode the audio latent and plug it into CreateVideo's optional audio input.
First prompt pays the load. The first clip of the day took 252 s; later clips of the same size took 183 to 232 s. Budget a minute or more for weights to stage into memory after a restart.

Settings

Two-stage, the way the Comfy LTX 2.5 templates do it. Stage one samples at half resolution (960x544) with 8 fixed sigmas (1.0, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875, 0.0); stage two runs the 2x latent spatial upscaler and then 3 more sigmas (0.85, 0.725, 0.4219, 0.0) at full 1920x1088. Sampler euler_ancestral, LTXV dual-CFG guider with video_cfg 1.0 and audio_cfg 1.0, frame rate 24. Stage-one seed per shot (listed under each clip), stage-two seed fixed at 42. Tiled VAE decode (tile 512, overlap 64, temporal 64/16). H.264 out of ComfyUI's SaveVideo node.

The i2v variant of the same graph (not shown) feeds the still into both stages with LTXVImgToVideoInplace at strength 1.0 after an LTXVPreprocess at compression 18. The workflow file linked below is the exact API-format JSON for the dachshund-b clip.

Speed and memory

Wall time is the batch script's clock from submitting the prompt to ComfyUI reporting it done, polled every 3 s. It covers both sampling stages, the latent upscale, the VAE decode and the mp4 encode, with the weights already resident.

Clip length	Clips	Wall time	Compute per second of video
121 frames (5.04 s) t2v	2	183.2 s, 186.2 s	36.6 s
97 frames (4.04 s) t2v	7	144.1 to 147.2 s	35.9 s
73 frames (3.04 s) t2v	1	114.1 s	37.5 s
121 frames (5.04 s) i2v	4	200.1 to 232.3 s	42.3 s
97 frames (4.04 s) i2v	1	153.4 s	38.0 s

So call it 36 seconds of compute per second of 1920x1088 text-to-video on the GB10, and roughly 42 for image-to-video. Time scales almost linearly with frame count. We haven't measured a lower resolution with this exact graph, so no number for 1280x704 yet; that's the next test when the GPU is free.

Memory: the ComfyUI process that holds these weights shows 38.5 GB in nvidia-smi after the run, which matches what the server log staged (text encoder 14.6 GB, transformer 20.5 GB, upscaler 0.9 GB, VAEs 2.1 GB). The GB10's memory is unified, so that figure is this process's share, not a system total. Peak during sampling was not measured; it was at least that, and it fit alongside a 19 GB voice model and a second ComfyUI without anything being evicted.

ComfyUI workflow for the dachshund-b clip (API format) ↗