Skip to content

ATTRIBUTION.md

Latest commit

 

History

History
61 lines (43 loc) · 4.37 KB

ATTRIBUTION.md

File metadata and controls

61 lines (43 loc) · 4.37 KB
Raw

Attribution

aeon-movie-maker orchestrates the LTX 2.3 video pipeline. Full credit to:

Models

LTX 2.3 22B (distilled fp8 + non-distilled fp8)

VBVR physics LoRA

  • The Verifier-Based Video Refinement (VBVR) physics LoRA improves motion realism in the LTX fast pipeline. Sourced from the ltxv community on HuggingFace.

Gemma abliterated encoder (text encoder)

  • Quantized GGUF (Q6_K) of an abliterated gemma-3-12b-it variant, used as the LTX text encoder.
  • HuggingFace: search for gemma-3-12b-it-abliterated.

IC-LoRA Union (optional)

  • Composition control via reference frames. Used for character continuity when seed-offset alone isn't enough.

Style LoRAs (Civitai-hosted, optional)

The following LoRAs are referenced by style: tags in screenplay JSON. They are hosted on Civitai and require a Civitai API token (CIVITAI_TOKEN in .env) for download. All are optional — plain prompts without style: tags don't need any of them.

Style tag LoRA filename Source Description
style: cyberpunk CyberPunkAI.safetensors Civitai community LoRA Neon / tech noir aesthetic
style: tribal Smooth_Tribal.safetensors Civitai community LoRA Ornamental / pattern-rich
style: claymation Claymation.safetensors Civitai community LoRA, LTX-targeted Stop-motion / clay
style: ghibli StudioGhibli.Redmond-StdGBRRedmAF-StudioGhibli.safetensors Civitai community LoRA Ghibli watercolor
style: ghibli_offset ghibli_style_offset.safetensors Civitai community LoRA Lighter Ghibli shift
style: galaxy LTX23-GalaxyAce.safetensors Civitai community LoRA, LTX 2.3-targeted Cosmic / nebula / starfield
style: illustration Illustration concept Variant 3A.safetensors Civitai community LoRA Illustrative / graphic

Search Civitai by filename to find each model's page; the API download URL pattern is https://civitai.com/api/download/models/<version_id> with Authorization: Bearer $CIVITAI_TOKEN. License terms are set per-LoRA by the original Civitai uploader — check each model's page before redistribution. This repo does not bundle or redistribute these LoRA weights — users must download them themselves under the original LoRA license terms.

ComfyUI custom nodes

Most LTX 2.3 nodes are in mainline ComfyUI. The full set:

  • LTXAVTextEncoderLoader, LTXAVDualCLIPEncode, LTXVImgToVideoInplace, LTXVBaseSampler, LTXVEmptyLatentAudio, LTXVSeparateAVLatent, LTXVReferenceAudio, LTXVImageEncoderForA2V
  • Standard: UNETLoader, VAELoader, LoraLoader, KSamplerSelect, BasicScheduler, RandomNoise, SamplerCustomAdvanced, SaveVideo

Audio mix chain

The stitch subcommand uses FFmpeg for sidechain ducking + final loudnorm. Filter chain documented in the aeon-radio-drama repo's references/mixing-guide.md.

Pipeline-specific design notes

  • Last-frame carry-forward: between sequential clips in a screenplay, the final frame of clip N is exported as a JPG and re-injected as the seed image for clip N+1, preserving character + lighting + composition continuity.
  • Persistence knob: scales the i2v denoise strength inversely (0 = full transformation, 1 = locked to seed). Original to this project.
  • Per-character seed offsets: stable hash from character name → small integer added to the base seed. Same character = same hash = same offset = same visual identity across an entire screenplay.
  • Per-scene LoRA routing: maps style_tags in screenplay JSON to LoRA file selection at workflow-build time.

License notes

This repo is MIT-licensed. Models retain their own licenses — refer to: