MiniMax H3 • Long Media Sampler

This is the heart of the pack. Everything else - Setup, Planner, the AV stream utilities - exists to feed this node. Long Media Sampler takes a long-media plan and expands it into a sequential multi-pass sampling graph: each segment is sampled in turn, context carried across the boundaries, and the results stitched into one final AV latent. When your manual duration exceeds your segment size, this is the node doing the heavy lifting, and it's also where every low-VRAM trick in the README actually lives.

The inputs that matter

Required: initial_av, long_media_plan, guider, sampler, sigmas, seed. Then a long list of knobs, but you only need to think about a handful:

• offload_completed_segments (default on) - moves each finished segment's stitched result to CPU RAM instead of piling the whole growing clip on the GPU. Pure VRAM win, zero output difference. Leave it on.
• video_context_denoise / audio_context_denoise - 0 preserves the inherited overlap exactly, 1 fully denoises it. The safe default is 0; nudge up only if a boundary looks stale.
• mlp_chunk_tokens (default 8192) - token chunk size for the low-VRAM MLP path. Bigger = faster but more VRAM; 0 disables chunking for A/B tests.
• attention_mode (default auto) - the big one. auto picks existing/Sage attention for small sequences and the embedded Sol path for long ones. existing forces stock attention; sol / scheduled_sol force the embedded Apache-2.0 SM120 Sol path.
• refine_enabled (default on) with refine_steps (default 2) - splits your SIGMAS into a main pass plus a low-noise refine tail. The README calls refine-on-with-2-steps the recommended production config. refine_add_noise and refine_seed are legacy stubs that get ignored.
• latent_hires_enabled (default off) - the optional learned H3 latent hi-res stage: upscale the video latent with a model from ComfyUI/models/latent_upscale_models, then run the refine pass at the higher resolution. Audio is preserved exactly. Set latent_hires_align to 32 to avoid edge artifacts.

The long VRAM guard family (vram_activation_reserve_mb, inter_block_vram_guard_mb, late_block_guard_*, step_boundary_cleanup_mb) are the emergency brakes. Leave them at defaults unless the console tells you otherwise.

The Sol attention story

Sol is embedded in the pack (an adapted subset of Saganaki22/ComfyUI-sol-attn under Apache-2.0) - no separate install. On very long sequences it can stream QKV in token chunks (retaining K/V as INT8+scale), compress K/V into block summaries, and chunk the output projection. That's the difference between a 30-second single pass being theoretically possible and it OOMing at the first attention layer. The sol_* knobs (tau schedule, min tokens, dense percent) are for when you want to tune that approximation; for most people auto is the right answer and you never touch them.

Outputs

final_av (the stitched latent, straight into Long Media Decode), total_frames, trim_frames, passes (how many segments actually ran - useful to confirm segmentation behaved), and a report.

Install & config

Pack install is the same as ever:

cd ComfyUI/custom_nodes
git clone https://github.com/vizart-vj/ComfyUI-MiniMax-H3-LongMedia

restart, then load an H3 checkpoint + video/audio VAEs (~42.5 GB from MiniMaxAI/MiniMax-H3; community license excludes US/EU/UK/Korea). The README's production advice: sampler_mode=auto, Dynamic VRAM on, and don't launch with --disable-dynamic-vram. For segment sizing, expect roughly 7–10 s on 16 GB, 5–8 s on 12 GB, 4–6 s on 8 GB with a low reference budget.

Troubleshooting

The community reality of H3 is that it's a big model - people on 24 GB cards are fine, and anyone below that is trading speed for VRAM via exactly these knobs. If you OOM, drop segment_seconds in Setup, set reference_budget=low, and confirm offload_completed_segments is on before touching the exotic sol knobs. If the first step of a long run dies specifically in attention, lower mlp_chunk_tokens and sol_qkv_chunk_tokens toward 4096 or 2048. And if seams appear at clip boundaries, check video_context_denoise - at 0 the inherited overlap is preserved exactly, and that's usually what fixes a jump cut.