ApolloRaines/LTX-2.5-22b-OmniGen-v12

LTX-2.5-22b OmniGen v12

A permanent weight merge of all five Lightricks IC-LoRA adapters into a single LTX-2.5 22B distilled model, plus a spectral SVD edit for temporal stability. No runtime adapters needed -- every LoRA's learned weight delta is baked into the base weights.

Modified from Lightricks/LTX-2.5: merged five IC-LoRAs at controlled scales, then applied a spectral SVD edit to attention and feedforward output weights. This derivative is distributed under the LTX-2.x Community License Agreement.

Format: GGUF Q6_K (18.6 GB), ready to drop into ComfyUI with ComfyUI-GGUF.

What This Is

Lightricks released five separate IC-LoRA adapters for LTX-2.5, each adding a distinct image-to-video capability. The adapters are designed for separate use, not simultaneous stacking. This model merges all five permanently into the base weights at controlled scales, then applies a spectral edit targeting temporal coherence.

The result is a single model file with all five LoRA weight deltas baked in. Capability-specific behavior (HDR, alpha, restore, etc.) still requires the appropriate reference input, as the original IC-LoRAs are in-context adapters. What the merge provides is improved temporal stability on image-to-video generation, measured on hard multi-subject and fine-detail shots.

Merged IC-LoRAs

LoRA	Scale	Source
Refine Details 1.0	0.60	Lightricks/LTX-2.5-22b-IC-LoRA-Refine-Details
Restore 1.0	0.30	Lightricks/LTX-2.5-22b-IC-LoRA-Restore
SDR-to-HDR 1.0	0.30	Lightricks/LTX-2.5-22b-IC-LoRA-SDR-To-HDR
Alpha Gen 0.9	0.15	Lightricks/LTX-2.5-22b-IC-LoRA-Alpha-Gen
Layout-to-Render 1.0	0.30	Lightricks/LTX-2.5-22b-IC-LoRA-Layout-To-Render

Merge formula per layer: W_new = W_base + (alpha / rank) * scale * B @ A

Total: 3,072 weight matrices modified across all five LoRAs.

Spectral Edit

After the LoRA merge, I applied an SVD boost to the top-64 singular values on 144 weight matrices (3 targets x 48 transformer layers):

Target	Boost
attn2.to_out (cross-attention output)	+0.25
attn1.to_out (self-attention output)	+0.10
ff.net.2 (FF down projection)	+0.15

Boost is scaled by layer position: early layers (0-15) get 20-30%, middle (16-31) get 50-70%, late (32-47) get 80-100% of the listed values.

Results

Single seed per shot. Treat as preliminary -- multi-seed confirmation in progress.

Image-to-Video (8 shots: 4 scenes x 2 styles, seed-matched against baseline)

Shot	flick/grad Change	Detail (grad)	Verdict
squad_photo	-18%	+6%	Win
clock_photo	-27%	+7%	Win
squad_anime	-12%	-1%	Win
hero (both), clock_anime, turbine (both)	+/-3%	+/-1%	Tie

flick/grad = flicker normalized by edge detail. Lower is more stable without losing sharpness.

3 of 8 shots improved, 5 unchanged. The biggest visible win: baseline duplicates and melts characters during fly-past sequences. v12 maintains coherent identity until the actual moment of passing.

What's Not Fixed

• Clock numerals and text/glyphs still degrade during zooms (both baseline and v12)
• Fast fly-past (object fills frame in a few frames) still smears
• Part of the flicker reduction comes with less motion (-9% to -17% on the winning shots)
• Capability-specific IC-LoRA behavior (HDR, alpha, restore, layout-to-render) has not been tested with reference inputs

Usage

Drop the GGUF into your ComfyUI models/unet/ folder. Use with:

• UNET Loader: UnetLoaderGGUF node
• Text Encoder: gemma4-12b-with-proj-ltx-2.5 (any quantization)
• VAE: ltx-2.5-video-vae-bf16.safetensors
• Sampler: euler, 8 steps, cfg 1.0
• Model Sampling: ModelSamplingLTXV with max_shift=2.05, base_shift=0.95

Works for both text-to-video and image-to-video at 1280x704, 97 frames (24fps).

Pipeline

BF16 safetensors (base)
  -> merge 5 IC-LoRAs at controlled scales
  -> spectral SVD edit on attention + FF outputs
  -> convert to BF16 GGUF
  -> quantize to Q6_K
  -> binary merge with original GGUF (preserves keyframes tensor)

License

This model is a derivative of Lightricks/LTX-2.5 and is distributed under the LTX-2.x Community License Agreement. A complete copy of the license is included in this repository.

Credits

Built on Lightricks/LTX-2.5. All five IC-LoRAs are Lightricks' work -- I found a way to combine them permanently and tuned the merge scales so they don't fight each other.

Independent derivative; not affiliated with or endorsed by Lightricks.

Model by Apollo Raines.