LTX-2.5 22B IC-LoRA Pixel Spatial Upscaler
This is a Pixel Spatial Upscaler IC-LoRA trained on top of LTX-2.5-22B, currently available in a 2x variant. It creatively upscales a low-resolution video, synthesizing fine detail rather than simply interpolating — making it a generative upscaler rather than a refiner.
It is based on the LTX-2.5 foundation model.
Model Files
ltx-2.5-22b-ic-lora-pixel-spatial-upscaler-x2-1.0.safetensors— the 2x upscaler checkpoint.
Model Details
- Base Model: LTX-2.5-22B Video
- Training Type: IC-LoRA (video-to-video, reference-conditioned)
- Control Type: Low-resolution reference video drives a 2x high-resolution re-render with synthesized spatial detail.
- Adapter Rank: 32
- Precision: BF16
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.
Control Signal Requirements
- Control signal: a low-resolution video clip (the reference).
- Reference downscale factor:
2. The reference is expected at half the linear resolution of the target output; the model's metadata carriesreference_downscale_factor: 2so compatible tooling can derive this automatically. - Framing: the reference must cover the same shot, duration and framing as the intended output. This is not an inpainting or reframing model — changing the aspect ratio or crop between reference and target degrades results.
- Source quality: the reference should be a clean low-resolution render, not a heavily compressed or degraded clip. See Intended Use — artifact removal is out of scope.
Usage
🔌 ComfyUI
- Copy the LoRA weights into
models/loras. - Load the LTX-2.5-22B base model and add
ltx-2.5-22b-ic-lora-pixel-spatial-upscaler-x2-1.0.safetensorsas the LoRA. - Use an IC-LoRA (video-to-video) workflow from the LTX-2 ComfyUI repository. Connect your low-resolution clip as the reference video.
Recommended Settings
- LoRA strength:
1.0. The weights ship pre-scaled so that full strength is the intended default. - Reference downscale factor:
2(see Control Signal Requirements). - Recommended flow: generate your video at a low base resolution (~280p) until you are satisfied with composition and subject motion, then feed that clip as the reference and run the 2x upscaler to produce the final high-resolution output.
The level of fidelity to the original can be traded off via generation parameters (guidance and step count in particular). Fewer steps and lower guidance keep the output closer to the reference; more steps and higher guidance allow more creative hallucination of detail.
Intended Use
Intended use: Creative upsampling as part of a generation flow. The recommended workflow is to generate a draft at a very low base resolution (e.g. ~280p) to nail composition and motion, then run the upscaler to produce a high-resolution result with synthesized detail.
Out of scope: This is not a blind denoiser or a compression-artefact remover. The model synthesizes new detail rather than faithfully preserving the reference — use it as a creative step, not a pixel-accurate refiner. It is also not intended for upscaling live-action footage where factual fidelity to the source matters.
References
- Code: GitHub Repository
- ComfyUI: ComfyUI-LTXVideo
- IC-LoRA docs: IC-LoRA usage guide
License
See the LTX-2-community-license for full terms.
Acknowledgments
- Base model by Lightricks
- Training infrastructure: LTX-2 Community Trainer