How we tested
Three prompts from our LTX 2.5 product-shot batch, rendered again with LTX 2.3 on the same Spark. Same text, same seeds, same 1920x1088 size, same frame counts, same two-stage sigma schedule. The only things we swapped were the model files and the sampler that each model's own ComfyUI template uses. The LTX 2.5 clips on this page are the ones from our [LTX 2.5 review](/video-lab/ltx-2-5-22b-distilled-text-to-video).
We picked those three because each one asks for something different:
- The dachshund asks for one small move: turn the head to camera and blink. LTX 2.5 nailed it, so it's the bar.
- The pantry asks for a slow sideways camera glide. LTX 2.5 did that one too.
- The freezer asks for the big move: a drawer sliding open toward the camera. LTX 2.5 ignored it.
To make the speed comparison fair, we also re-ran LTX 2.5 on two of the prompts the same morning, on the same idle machine. Those re-runs came out frame-for-frame identical to the October clips, so the clips on the page and today's timings belong together.
The clips are above, each pair in a side-by-side player. One thing the stills can't show: LTX 2.3's drawer move lurches a little. The drawer front drops out of the bottom of the frame as the camera seems to dive in. The move still happened, and on a product shot that's the part you can't fake.
Setup and gotchas
- Nothing new to download if 2.3 is already on your box. We ran it in the same ComfyUI 0.43.0 as LTX 2.5, with no custom nodes. The graph follows ComfyUI's own LTX-2.3 text-to-video template: the 2.3 checkpoint (it carries the video VAE and audio VAE inside it), the distilled LoRA on top at strength 0.5, the Gemma 3 12B text encoder, and the 2.3 x2 latent upscaler.
- Our distilled LoRA is the older one. The template names version 1.1 of the distilled LoRA. We had the original
ltx-2.3-22b-distilled-lora-384on disk and used it at the template's 0.5. It worked fine, but if you're starting fresh, get 1.1. - Starting from scratch, 2.3 is the bigger download. The 2.3 set is about 47 GB: the fp8 dev checkpoint (29.1 GB), the distilled LoRA (7.6 GB), Gemma 3 12B fp4 (9.4 GB) and the upscaler (1.0 GB). LTX 2.5's set is about 39 GB.
- Different text encoders. LTX 2.3 reads prompts with Gemma 3 12B, LTX 2.5 with Gemma 4 12B. You can't share one file between them.
- 2.3's template decodes sound; our 2.5 graph didn't. The 2.3 clips came out with an audio track. We stripped it on this page so you're comparing pictures with pictures.
- The negative prompt does nothing on either. Both models run at CFG 1, so the negative prompt has no effect either way.
Settings
Text-to-video, 1920x1088, 24 fps. The dog is 121 frames (5.04 s); the pantry and freezer are 97 frames (4.04 s). Two stages: stage one samples at 960x544 with 8 fixed sigmas (1.0, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875, 0.0), then the x2 latent upscaler, then 3 more sigmas (0.85, 0.725, 0.4219, 0.0) at full size. Stage-one seeds 2207 (dog), 909 (pantry), 808 (freezer); stage-two seed 42. Tiled VAE decode (tile 512, overlap 64, temporal 64 / 16).
| LTX 2.3 | LTX 2.5 | |
|---|---|---|
| Model file | ltx-2.3-22b-dev-fp8 + distilled LoRA 384 at 0.5 | ltx-2.5-22b-distilled-transformer-comfy-int8-convrot |
| Text encoder | Gemma 3 12B fp4 mixed | Gemma 4 12B int8 |
| Sampler and guider | euler, CFG guider 1.0 | euler ancestral, LTXV dual-CFG guider 1.0 / 1.0 |
| Upscaler | ltx-2.3-spatial-upscaler-x2-1.1 | ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0 |
The workflow file linked below is the exact API-format graph we sent for the 2.3 pantry clip.
Speed and memory
Wall time is from sending the graph to ComfyUI until it reported the clip done. It covers both sampling stages, the latent upscale, the VAE decode and the mp4 write. "Cold" means we had just told ComfyUI to unload everything, so the clip paid for loading the weights.
| Clip | Model | Wall time | Compute per second of video |
|---|---|---|---|
| Pantry, 4.04 s, warm | LTX 2.3 | 91.3 s | 22.6 s |
| Freezer, 4.04 s, warm | LTX 2.3 | 91.2 s | 22.6 s |
| Dog, 5.04 s, cold | LTX 2.3 | 149.9 s | 29.7 s |
| Freezer, 4.04 s, warm | LTX 2.5 | 139.8 s | 34.6 s |
| Pantry, 4.04 s, cold | LTX 2.5 | 163.0 s | 40.3 s |
Warm, LTX 2.3 finished the same 4-second clip in 91 s against 140 s for LTX 2.5. That's about 35% less time, on the same machine, the same morning, with nothing else on the GPU. We didn't break down where the difference comes from (fp8 checkpoint versus int8 transformer, the text encoder, the sampler), so take it as the speed of each setup as shipped.
Memory came out the same. The ComfyUI process peaked at 40.5 GiB with LTX 2.3 and 40.6 GiB with LTX 2.5, read from nvidia-smi once a second. On the GB10's unified memory that's this process's share, not a machine total. The lowest free memory we saw on the box during a run was 51 GiB with 2.3 and 46 GiB with 2.5.
What to do with this
If LTX 2.3 is already on your Spark, keep it. Reach for it when you want the shot to look like footage, when a prop has to move, or when you're burning through seeds and time matters. Reach for LTX 2.5 when the subject has to look clean and stay put, or when the move is a camera glide. Either way, don't trust either one with readable text: put labels in afterwards.