gpu-container-ltx

RunPod serverless container for LTX 2.3 video generation via ComfyUI.

Based on Jef's workflow — two-pass I2V with LCM sampler, spatial upscaling, and audio generation. Runs the full (dev) transformer + distill LoRA so the camera-control LoRAs bind while keeping fast few-step LCM sampling.

Models Included

Model	Size	Source	Purpose
ltx-2.3-22b-dev_transformer_only_fp8_scaled.safetensors	~11GB	Kijai/LTX2.3_comfy	Full (dev) transformer (FP8) — camera LoRAs bind to this
ltx-2-19b-distilled-lora-384.safetensors	~?GB	Lightricks/LTX-2	Distill LoRA @ 0.6 — restores few-step LCM on the dev transformer
gemma_3_12B_it_fpmixed.safetensors	~6GB	Comfy-Org/ltx-2	Text encoder (mixed precision)
ltx-2.3_text_projection_bf16.safetensors	~2.3GB	Kijai/LTX2.3_comfy	Text projection for DualCLIPLoader
LTX23_video_vae_bf16.safetensors	~1.5GB	Kijai/LTX2.3_comfy	Video VAE
LTX23_audio_vae_bf16.safetensors	~365MB	Kijai/LTX2.3_comfy	Audio VAE
ltx-2.3-spatial-upscaler-x2-{1.0,1.1}.safetensors	~1GB each	Lightricks/LTX-2.3	2x latent upscaler (v1.0 + v1.1 hotfix)
taeltx2_3{,_wide}.safetensors	~50MB each	madebyollin/taehv	TAESD for fast preview during render
7x camera control LoRAs	~100MB each	Lightricks	Dolly in/out/left/right, jib up/down, static

Pipeline (from Jef's workflow)

Pass 1 — Low-res generation (704x512, 121 frames):
  UNETLoader → LoraLoaderModelOnly (distill LoRA @ 0.6) → Power Lora Loader (camera LoRA @ 0.4)
  DualCLIPLoader → CLIPTextEncode → LTXVConditioning
  LoadImage → LTXVPreprocess → LTXVImgToVideoInplace
  EmptyLTXVLatentVideo + LTXVEmptyLatentAudio → LTXVConcatAVLatent
  → SamplerCustomAdvanced (LCM, 8 steps, LTXVScheduler)
  → LTXVSeparateAVLatent

Pass 2 — Upscaled refinement:
  → LTXVLatentUpsampler (2x spatial)
  → LTXVImgToVideoInplace (re-inject ref image)
  → LTXVConcatAVLatent (recombine with audio)
  → SamplerCustomAdvanced (LCM, 3 steps via ManualSigmas: 0.909, 0.725, 0.422, 0.0)
  → LTXVSeparateAVLatent

Output:
  Video → VAEDecodeTiled (tile=512, overlap=64)
  Audio → LTXVAudioVAEDecode
  → VHS_VideoCombine (H.264 MP4, CRF 19, 24fps)

Custom Nodes

• ComfyUI-LTXVideo — LTX-specific nodes
• ComfyUI-VideoHelperSuite — VHS_VideoCombine
• ComfyUI-KJNodes — VAELoaderKJ, SimpleCalculatorKJ, ImageResizeKJv2, SetNode/GetNode
• rgthree-comfy — Power Lora Loader (multi-LoRA support)

Key Settings

• Sampler: lcm (both passes)
• CFG: 1.0
• Pass 1 scheduler: LTXVScheduler (steps=8, max_shift=2.05, min_shift=0.95)
• Pass 2 scheduler: ManualSigmas (0.909375, 0.725, 0.421875, 0.0)
• Distill LoRA: ltx-2-19b-distilled-lora-384 @ 0.6 (enables few-step LCM on dev model)
• Camera LoRA: static @ 0.4 strength (default)
• Output: 24fps, H.264, CRF 19

Requirements

• NVIDIA GPU with 32GB+ VRAM (L40S, A100 80GB, H100 recommended)
• ~60GB disk for models + cache

Build

docker build -t edream/gpu-container-ltx:latest --target final .

Local Dev

docker-compose up  # ComfyUI on :8188, RunPod handler on :8000

Algorithm

infinidream_algorithm: "ltx-i2v"