H3 Tiled Sampler
A dedicated tiled (block-based) sampling node for the H3 video model, using LTX 2.3-style 2D (H x W) tiled sampling.
Design goals (inspired by LTX 2.3 tiling, adapted for H3's dual-modality setup):
- Everything stays on VRAM throughout; data is never dropped to CPU memory (the main cause of near-freezes / slow generation).
- Tile count/size auto-adapts to available VRAM, so the user doesn't have to guess n_tiles, and a bad guess won't cause an OOM.
- Reuses the global noise + global text conditioning (guider.raw_conds), slicing only the latent into tiles -> preserves the prompt as much as possible and reduces drift.
- Each tile is freed right after it's computed; no torch.cuda.empty_cache() inside the loop (that forces a sync and slows things down) -- only a single soft_empty_cache() at the very end.
- Supports 2D H x W tiling with separable cosine-window blending (outer product of two 1D windows) to remove seams.
Key differences from LTX 2.3:
- H3 outputs a (video, audio) tuple with two modalities; audio passes through untouched and does not participate in sampling.
- Duration is determined by VAE encoding; the entry point only applies a minimal frame-count safety floor and never truncates.
- Does not assume the model has process_latent_out -- detected via hasattr.
Intended use case: H3 768p/2K upscale-refine passes, where too many tokens hurt attention performance -- spatial tiling keeps the token count per tile within the range the H3 DiT was trained on. Not suitable for: full re-denoising from pure noise. Spatial tiling breaks global consistency in that case.
This node based on LTX Tiled Sampler of TenStrip for LTX 2.3 (https://github.com/TenStrip/10S-Comfy-nodes)
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support