Loading
Du hast einen ComfyUI-Workflow? Erstelle direkt im Browser eine verständliche Workflow-Dokumentation.Workflow verständlich machen.Jetzt ausprobieren ↗
Everything here is free and stays free — the format spec, the nodes, the workflows, the cartridges, the LoRAs. If it saved you a night of debugging (it contains several hundred of mine), tips keep the 5090 warm:
Modellquelle
Quellenbeschreibung
For 5090 and workstation cards. Two builds, one per loader.
LTX25-distilled-DiT-Q8_0.gguf - 22.73 GB . Needs ComfyUI-GGUF . Top of the Q-ladder; effectively indistinguishable from the bf16 original.
LTX25-distilled-DiT-comfy-int8.safetensors - 21.50 GB . No custom node - stock Load Diffusion Model on ComfyUI 0.32+. Per-channel scales with the convrot rotation, the same format Lightricks ship themselves, and the widest kernel support of anything in this listing: INT8 tensor cores go back to Turing. If a 4-bit or FP8 build ever falls back to something slow on your card, this is the one that will not.
Why there is no FP8 build here
There was, and it was cut on measurements rather than taste. Across 24 identical real layers, int8_tensorwise reconstructs at 0.00942 relative error against float8_e4m3fn at 0.02682, mxfp8 at 0.02656 and float8_e5m2 at 0.05289 - all at the same 21.5 GB except mxfp8, which is 0.6 GB larger.
Quellen
1 QuelleModellartefakte
2 Artefakteltx2522BCutDownToFit_32gbGpus.safetensors
safetensors · INT8 · 20,0 GB · SHA-256 a9adb1c33604…0326 · Civitai
HerunterladenQuellenauszüge
1 AuszugSource context: 472 downloads · Type Checkpoint · Base model LTXV 2.5
Speed was the last argument for keeping FP8, and it does not survive either. Timed on an RTX 5090 at 4096 tokens, milliseconds per forward pass: int8 0.264 , 4-bit w4a8 0.287, mxfp8 0.228, float8_e4m3fn 0.415 - that is 1.58x SLOWER than int8 while also being three times less accurate. float8_e5m2 does not run at all: PyTorch refuses to multiply two e5m2 matrices.
So int8 is the 8-bit build, and there is no card or workload for which the FP8 ones are the better choice.
Read this before you take the biggest file
At this tier you are buying certainty, not visible quality . On identical scene, seed and size, nothing above Q4_K_M separates cleanly by eye. What a 22 GB file costs you is the activation pool , and on a 32 GB card that arithmetic gets tight faster than people expect: 22.73 GB of weights on a 32 GiB card leaves roughly 8 GB for activations, and a long high-resolution render will want more.
The failure is not an out-of-memory error. On Windows the driver falls back to shared system memory silently and the render simply crawls - GPU pinned at 100%, wattage low, seconds-per-step up by more than an order of magnitude. Measured here: 329 s/it where the same shape normally runs about 15 . If you see that, you have spilled, and the fix is a smaller quant rather than patience.
So take Q8_0 or comfy-int8 for short clips and stills work where you want no doubt about the weights. For long multi-shot renders, Q5_K_M or the mix4x8-17GB build from the 24 GB version will finish while these are still spilling.
One more thing that costs 32 GB owners more than anyone: a second idle ComfyUI does not give its VRAM back. An instance sitting with an empty queue held about 11 GB here and starved the one that was actually rendering down to a fifth of its normal wattage. If you keep two, free or stop the idle one first.
Also grab
gemma4-12b-ltx25-comfy-w4a8.safetensors from the 16 GB version - 10.60 GB . The bf16 encoder is 26.3 GB, so even here you want this or Lightricks' int8 build unless you are happy loading and unloading 26 GB per prompt.
