infosave

/

LTX-2.5-cmf

like 2

Thanks — implemented it. Shipped in 0.5.90, and 0.5.91 makes it cheap: cortiq ltx-video --lora … --ref ….

Your example 03 renders correctly on an M4 through the CMF engine, no ComfyUI and no torch: 480 branches of rank 128 bind, three references add 1152 tokens beside 384 of clip at frames −3, −2, −1, and all three slots land.

One finding worth passing on: the cost is placement, not flops. The branch is only 4% more multiply-adds, but where it runs decides everything — 182.7 s a step as scalar loops, 39.7 through a device probe, 22.2 pinned to the host, and 10.1 fused into the base GEMM's own Metal submission, against 8.6 with no adapter at all. In 0.5.91 the branch reads the activation that GEMM already uploaded and accumulates into the output it already wrote, so an adapter costs no transfer: 1.17× a step, not 2.6×.

Any LTX-2.5 adapter should load — --lora file.safetensors --lora-strength 0.8. If one refuses, tell me: it refuses loudly rather than rendering something plausible.

you work so fast

so this allows for run time integration of any lora for ltx?

can we do same with mini max?

TX-2.5 22B IC-LoRA Pixel Spatial Upscaler

https://huggingface.co/Lightricks/LTX-2.5-22b-IC-LoRA-Pixel-Spatial-Upscaler

How It Works

The model is an in-context LoRA: instead of taking only a text prompt, it also consumes a reference video in context alongside the noisy latents being denoised. During training, the reference is a downscaled version of the target clip, so the model learns to map a low-resolution reference onto a higher-resolution rendering of the same scene.

At inference the reference is your low-resolution clip. The model does not upsample it pixel-wise; it re-renders the scene at the higher resolution, inventing plausible fine detail (texture, edges, micro-contrast) that was never present in the reference. Composition, motion and subject identity are carried over from the reference, while high-frequency detail is generated.

so this allows for run time integration of any lora for ltx? can we do same with mini max?

Any LTX adapter whose branches land on the ported projections, yes — and 0.5.92 widens what "lands" means: the loader now accepts the three naming conventions in the wild (diffusion_model.…, base_model.model.…, and the bare module path) instead of one, so PEFT-trained adapters bind too. It also refuses, by name, a branch whose shape does not match the projection it would ride, instead of writing past the panel.

And yes, MiniMax too, as of 0.5.92 — cortiq animate --lora, with the details and the measured cost in the H3 thread.

Two new switches on both models: CMF_LORA_PROBE=1 prints every branch by its measured contribution ‖s·ΔY‖/‖Y‖, and CMF_LORA_ROUTE=<r> switches off the ones below r. On the H3 Realism adapter that map is lopsided — the loudest branch is 150× the quietest and 41 of 104 branches carry the look.

On the IC-LoRA Pixel Spatial Upscaler: it needs in-context video conditioning — the low-res clip rides in the sequence next to the noisy latents. We have that for reference stills (--ref), not for clips, so the adapter would load and then be conditioned on nothing. Before that lands, the engine's own two-stage path does the same job natively:

cortiq ltx-video --model ltx25-q4tp.cmf --prompt "…" --two-stage --steps2 3

Half resolution, learned latent upscale, three refinement steps at full size.

0.5.93 for the H3 side of this thread: the latent upscaler is ported and gated (worst 6.7e-6 against the reference module), --upscale on cortiq animate. The LTX two-stage path is unchanged — --two-stage still does the same job natively for LTX.