Loop Forge
Loop Forge / Projects

9 October 2026

LTX 2.5 Creative Lab LoRAs, tested locally

Lightricks has released a set of official LoRAs for LTX 2.5 that add things the base model does not do on its own, from building a shot out of a reference sheet to colouring a black-and-white film. I ran 14 of them in ComfyUI on my own desktop, plus three tests that chain two LoRAs in one generation.

Each test has a frame from every clip in it, the exact prompt, the workflow that ran and the inputs it was given. 25 of the 31 tests worked, and the other 6 show where a LoRA reached its limit.

Watch on YouTube

31 prompts

Ingredients

Works

Character, prop and location

Works

Guitarist on a stage

Limit

Two characters from one sheet

Limit

The sheet comes through

Colorization

Works

Grayscale market

Works

A Bucket of Blood, 1959

Slow Motion Control

Works

Running through Tokyo rain

Works

Dove in flight

Cinemagraph

Works

Portrait, hair in the breeze

Works

Standing figure

Limit

Tree full of falling leaves

Clean Plate

Works

Woman with umbrella

Works

Couple on a canal, moving camera

Day-to-Night

Works

Park with fountain

Works

Canal at night, moving camera

Limit

Same park at strength 1.0

Water Simulation

Works

Water into a sandbox

Works

Wave over the steps

Decompression

Works

Heavily macroblocked market

Works

Superman, 1941

Deblur

Works

Clowns on a city street

Works

1956 Navy film, diploma handover

Restore

Works

A Trip Down Market Street, 1906

Refine Details

Works

Generated canal clip at 2x

SDR to HDR

Works

Forest creek, sun through mist

Alpha Gen

Works

Curly hair in the rain

Works

Portrait, hair in the wind

Layout to Render

Limit

Town square, camera arc

Chained LoRAs

Works

Clean Plate with Day-to-Night

Works

Colorization with Decompression

Limit

Ingredients with Slow Motion

Run it yourself

Every test page has a zip with its workflow, every input it loads and a README listing the models it needs. Unzip it, copy the inputs into ComfyUI/input, put the models in the folders shown below and drag the workflow onto ComfyUI.

The workflows follow the official Lightricks templates, one graph per kind of LoRA. The model loaders and every ComfyUI-LTXVideo node stay in view; only core ComfyUI plumbing is folded into subgraphs, with the prompt and settings on them. Every workflow sizes its output from what you load with one setting, Shorter side, and the video ones take their length and frame rate from the clip, so they run on your own footage as they are. Every kind of workflow here was run from its downloadable file. The clips on these pages came from the original test graphs, with the same settings. If a node shows up red, install or update ComfyUI-LTXVideo to the commit listed rather than rewiring the graph.

The Lightricks repositories on Hugging Face, the LTX 2.5 base model and every LoRA, are gated. Log in, open each repository page and accept its terms before downloading, or the links return an error.

I ran everything on an RTX 3060 with 12GB of VRAM, an i5-13400F and 32GB of RAM. Peak VRAM across the tests was 7.5 to 11.8 GiB and each clip took 3 to 36 minutes.

Software

Models

FilePut it inSizeTests
ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors gatedmodels/diffusion_models21.5 GB31
gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors gatedmodels/text_encoders15.4 GB30
ltx-2.5-audio-vae-bf16.safetensors gatedmodels/vae364.9 MB31
ltx-2.5-video-vae-bf16.safetensors gatedmodels/vae1.5 GB31
ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors gatedmodels/latent_upscale_models995.8 MB3
ltx-2.5-22b-ic-lora-alpha-gen-0.9.safetensors gatedmodels/loras1.3 GB2
ltx-2.5-22b-ic-lora-clean-plate-1.0.safetensors gatedmodels/loras327.3 MB3
ltx-2.5-22b-ic-lora-colorization-0.9.safetensors gatedmodels/loras906.1 MB3
ltx-2.5-22b-ic-lora-day-to-night-0.9.safetensors gatedmodels/loras327.3 MB4
ltx-2.5-22b-ic-lora-deblur-0.9.safetensors gatedmodels/loras906.1 MB2
ltx-2.5-22b-ic-lora-decompression-0.9.safetensors gatedmodels/loras906.1 MB3
ltx-2.5-22b-ic-lora-ingredients-0.9.safetensors gatedmodels/loras1.3 GB5
ltx-2.5-22b-ic-lora-layout-to-render-1.0.safetensors gatedmodels/loras1.3 GB1
ltx-2.5-22b-ic-lora-refine-details-1.0.safetensors gatedmodels/loras1.3 GB1
ltx-2.5-22b-ic-lora-restore-1.0.safetensors gatedmodels/loras1.7 GB1
ltx-2.5-22b-ic-lora-sdr-to-hdr-1.0.safetensors gatedmodels/loras1.3 GB1
ltx-2.5-22b-ic-lora-water-simulation-0.9.safetensors gatedmodels/loras906.1 MB2
ltx-2.5-22b-lora-cinemagraph-0.9.safetensors gatedmodels/loras201.5 MB3
ltx-2.5-22b-lora-slow-motion-control-1.0.safetensors gatedmodels/loras654.5 MB4
ltx-2.5-22b-ic-lora-sdr-to-hdr-scene-emb.safetensors gatedmodels/embeddings12.6 MB1

Inputs

Rooftop sheet, Ingredients sheet: character, camera and rooftop panels on black
Rooftop sheetIngredients sheet: character, camera and rooftop panels on blackDownload PNG0.7 MB
Stage sheet, Ingredients sheet: character, guitar and stage panels on black
Stage sheetIngredients sheet: character, guitar and stage panels on blackDownload PNG0.7 MB
Two-face sheet, Two face crops, a guitar and a jazz club, 2×2 on black
Two-face sheetTwo face crops, a guitar and a jazz club, 2×2 on blackDownload PNG0.7 MB
Grayscale market, A generated market clip, desaturated with FFmpeg
Grayscale marketA generated market clip, desaturated with FFmpegDownload MP43.7 MB
A Bucket of Blood, Roger Corman, 1959. Public domain, via Wikimedia Commons
A Bucket of BloodRoger Corman, 1959. Public domain, via Wikimedia CommonsDownload MP42.8 MB
Tokyo still, Start still, made with Qwen-Image 2.1
Tokyo stillStart still, made with Qwen-Image 2.1Download PNG2.9 MB
Dove still, Start still, made with Qwen-Image 2.1
Dove stillStart still, made with Qwen-Image 2.1Download PNG2.3 MB
Kate, Character photo
KateCharacter photoDownload PNG7.2 MB
Allie, Character photo
AllieCharacter photoDownload PNG0.5 MB
Autumn tree, First frame of a clip generated with LTX 2.5
Autumn treeFirst frame of a clip generated with LTX 2.5Download PNG0.6 MB
Umbrella clip, Generated with LTX 2.5, no LoRA. 129 frames at 25fps
Umbrella clipGenerated with LTX 2.5, no LoRA. 129 frames at 25fpsDownload MP41.0 MB
Canal clip, Generated with LTX 2.5, no LoRA
Canal clipGenerated with LTX 2.5, no LoRADownload MP41.6 MB
Park clip, Generated with LTX 2.5, no LoRA
Park clipGenerated with LTX 2.5, no LoRADownload MP41.9 MB
Sandbox clip, Generated with LTX 2.5, no LoRA
Sandbox clipGenerated with LTX 2.5, no LoRADownload MP410.5 MB
Stairs clip, Generated with LTX 2.5, no LoRA
Stairs clipGenerated with LTX 2.5, no LoRADownload MP40.8 MB
Blocky market, A generated market clip, re-encoded at 120 kbps
Blocky marketA generated market clip, re-encoded at 120 kbpsDownload MP473 KB
Superman, Fleischer Studios, 1941. Public domain, via the Internet Archive. Scaled to 896×672
SupermanFleischer Studios, 1941. Public domain, via the Internet Archive. Scaled to 896×672Download MP48.4 MB
Clowns walking, By Coverr on Pexels, Pexels License. First 161 frames, scaled to 1536×864
Clowns walkingBy Coverr on Pexels, Pexels License. First 161 frames, scaled to 1536×864Download MP44.4 MB
Japanese Bride School, US Navy, 1956, National Archives 75330. Public domain, via the Internet Archive
Japanese Bride SchoolUS Navy, 1956, National Archives 75330. Public domain, via the Internet ArchiveDownload MP42.7 MB
A Trip Down Market Street, Miles Brothers, 1906. Public domain, via the Internet Archive. 640×480
A Trip Down Market StreetMiles Brothers, 1906. Public domain, via the Internet Archive. 640×480Download MP42.2 MB
Diana Creek, By Extemporalist, CC0, via Wikimedia Commons
Diana CreekBy Extemporalist, CC0, via Wikimedia CommonsDownload MP420.8 MB
Tokyo run, The Slow Motion LoRA result from this project
Tokyo runThe Slow Motion LoRA result from this projectDownload MP46.2 MB
Kate portrait, The Cinemagraph LoRA result from this project
Kate portraitThe Cinemagraph LoRA result from this projectDownload MP43.7 MB
Clay layout, Rendered in Blender from primitives with clay shading
Clay layoutRendered in Blender from primitives with clay shadingDownload MP49.0 MB
Look still, Qwen-Image 2.1 edit of the layout's first frame, with a character photo
Look stillQwen-Image 2.1 edit of the layout's first frame, with a character photoDownload PNG4.2 MB
Grayscale, blocky market, A generated market clip, grayscale and at 120 kbps
Grayscale, blocky marketA generated market clip, grayscale and at 120 kbpsDownload MP471 KB

The model, and the machine

Base model
LTX 2.5 22B distilled, int8 (ltx-2.5-22b-distilled-transformer-comfy-int8-convrot)
Text encoder
gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot
LoRAs
14 official Lightricks LoRAs, of which 12 are IC-LoRAs guided by a video or an image
Clips
About 5 seconds each, mostly 121 frames at 24fps, at each LoRA's trained size or the largest size that worked
Render times
3 to 36 minutes per clip
Ran on
My RTX 3060 with 12GB, i5-13400F, 32GB RAM, PyTorch on CUDA 13

Materials and links

What we found

Which of the LTX 2.5 LoRAs are worth using on a 12GB card?

The video to video LoRAs were the most dependable. Clean Plate, Colorization, Decompression, Deblur, Restore and Alpha Gen worked on every clip I gave them, including real archive footage, and Day-to-Night worked once I raised its strength to 1.5 at 1152×640.

The ones that generate new motion needed more care. Cinemagraph worked on the two people I tried, but a tree full of leaves drifted across the whole frame. Ingredients kept one character from a reference sheet and lost the faces when I gave it two, and Layout to Render followed the layout but turned the character's back to the camera.

Chaining two LoRAs in one generation worked for Clean Plate with Day-to-Night and for Colorization with Decompression. Ingredients with Slow Motion failed on all three seeds I tried.