H3 Relay · LTX Upscale Model Loader

Raw MiniMax H3 comes out at native 480p. That's fine for review, but it's not the movie you want to show anyone. H3RelayLTXModelLoader - displayed as H3 Relay · LTX Upscale Model Loader - is where H3 Relay gets the machinery to fix that: it loads the entire LTX 2.5 finishing stack and bundles it onto the single H3_RELAY_MODEL wire that fans out to every LTX 2× Enhance shot node.

Here's the stack it assembles, and it's a real shopping list: the LTX 2.5 22B dev transformer (the int8-convrot Comfy build), the LTX video VAE, the learned latent 2x spatial upscaler, the Gemma4 12B text encoder, and two LoRAs - the distilled LoRA and the pixel-spatial IC-LoRA. Six files, one node.

Two LoRAs, two jobs - don't confuse them

The most common misunderstanding is treating the upscale as one thing. It's a one-2x-pipeline, two-stage affair:

1. latent_2x_model_name - the learned latent upscaler expands the target latent 2x. This establishes the high-resolution target; it does the resolution work.
2. pixel_upscale_ic_lora - a reference-conditioned IC-LoRA that then guides diffusion refinement from the original low-resolution pixel video, keeping detail honest to what H3 actually generated.

Meanwhile distilled_lora (the ltx-2.5-22b-distilled-lora-450-bf16) does a third, separate job: it adapts the dev transformer for the fast low-step schedule. The tooltip is explicit - it does not perform the spatial upscale. All three at strength 1.0, per the reference workflow.

The inputs that matter

• model_name, vae_name, latent_2x_model_name, text_encoder_name - dropdowns from their respective model folders. Defaults point at the exact files in MODELS.md.
• distilled_strength (default 1, −4 to 4) and pixel_upscale_ic_strength (default 1, 0 to 2) - the two LoRA weights.
• weight_dtype - default, or fp8_e4m3fn / fp8_e4m3fn_fast / fp8_e5m2. The default is the quality choice; reach for an fp8 variant when VRAM is tight, which on a 22B transformer plus a 12B encoder it usually is.
• manual_cache_revision (default v1) - the advanced reset key for replacing files without changing names; everything else is auto-tracked via fingerprint.

Output: ltx_model (H3_RELAY_MODEL) - one typed wire into every LTX 2× Enhance node. The bundle carries the model, VAE, latent upscaler, text encoder, required LoRAs, and an internal cache fingerprint, so a chain change invalidates only derived artifacts.

Install - the heavy part

The pack itself is light: ComfyUI Manager → H3 Relay, restart, or git clone https://github.com/akatz-ai/h3-relay into custom_nodes. The models are the heavy part and none of them ship with the pack:

hf download Lightricks/LTX-2.5 \
  diffusion_models/ltx-2.5-22b-dev-transformer-comfy-int8-convrot.safetensors \
  text_encoders/gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors \
  vae/ltx-2.5-video-vae-bf16.safetensors \
  latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors \
  loras/ltx-2.5-22b-distilled-lora-450-bf16.safetensors \
  --local-dir models

plus the pixel-spatial IC-LoRA from Lightricks/LTX-2.5-22b-IC-LoRA-Pixel-Spatial-Upscaler into models/loras. Empty dropdowns = missing files. And remember: these weights are separately licensed from the pack's GPL-3.0 source.