Lightricks

/

LTX-Video

like 2.25k

Follow

LTX.io 4.14k

Posting here since issue creation is restricted on the GitHub repos. Hoping someone from Lightricks or the community can sanity-check my approach.

Symptom

A custom character LoRA trained for ltxv-13b-0.9.7-dev.safetensors loads into ComfyUI with no errors and no warnings, but produces zero change in output. Sampling through LTXVBaseSampler returns generic, unconditioned subjects with no trace of the trained identity.

The important detail: this is not a weak or degraded likeness that might indicate undertraining. It is no effect at all. Sweeping LoRA strength from 0.0 to 2.0 produces byte-identical results at a fixed seed. That strongly suggests the adapter is being read from disk but never actually patched into the transformer at inference time.

Goal

Character-consistent I2V and T2V generation of a stylized 2D cel-animated character — flat color fills, hard black line art, limited-animation motion. Building toward long-form animated work, so consistency across separate generations matters more to me than per-clip fidelity.

Environment

Key naming. Remapped extracted tensor keys to model.diffusion_model.transformer_blocks.X... across all 48 transformer blocks to match the base checkpoint's structure. Names now align on inspection.

Strength. Swept 0.0 → 2.0. No change at any value, fixed seed. This is the finding I keep coming back to.

Checkpoints. Tested multiple step checkpoints across the training run. Same null result on all of them.

Environment. Dependencies current (transformers 5.14.1, current CUDA backends), VRAM caches flushed between runs, node registry verified, ComfyUI restarted clean. Nothing logged at load time.

Questions

1. Runtime verification. Is there a supported way to confirm a LoRA is actually patching the transformer rather than silently no-op'ing? A logging flag, a debug node, or a Python snippet that diffs weights pre- and post-patch would let me isolate whether this is a training problem or a loading problem. Right now I can't tell which half of the pipeline is at fault, and that's blocking everything else.

2. Trainer compatibility. Has anyone successfully applied a diffusion-pipe LoRA to 13B in ComfyUI? I'm aware of the official LTX-Video-Trainer and am willing to retrain there, but I'd like to know whether the diffusion-pipe path is expected to work at all before I spend the compute. If there's a canonical key-format or adapter-config spec I should be converting to, a pointer would be ideal.

3. Layer targeting. For character-identity LoRAs on 13B (vs. 2B), what layers are recommended — attention QKV and output projections, feed-forward, or both? And does STG interact with LoRA strength in a way that can suppress adapter influence?

4. Stylized 2D content. Are there known limitations or recommended hyperparameters (rank, alpha, LR, dataset size, clip length) for flat 2D cartoon assets? LTX's latent space appears tuned for photographic motion, and I'd like to understand the realistic ceiling before scaling up.

5. Version strategy. I'm on 0.9.7. I know 0.9.8 shipped with 60-second generation, and that LTX-2 is now the primary home for development with its own LoRA trainer and ComfyUI core integration. If character LoRAs are meaningfully better supported on LTX-2, tell me and I'll migrate now rather than debug a branch in maintenance mode.

Known limitations I should plan around

Separate from this bug — I'd genuinely value candor here. Max reliable clip duration and resolution, identity drift over long sequences, and the practical ceiling on character consistency across independent generations. I'd rather design my pipeline around real constraints than discover them at scale.

Happy to share training configs, the ComfyUI workflow JSON, or a minimal repro. Thanks for building this in the open.

— Keith