feat: support Sol-H3-VAE decoding for MiniMax-H3 (#2317)

docs/minimax_h3.md

@@ -99,6 +99,24 @@ The C API exposes the same inputs through `ref_images`, `ref_videos`, and
 `ref_audios` in `sd_vid_gen_params_t`. Each `sd_ref_video_t` supplies its own
 frame rate and optional soundtrack; non-24-fps inputs are resampled internally.
 
+## Sol-H3-VAE decoding
+
+Pass a [Sol-H3-VAE checkpoint](https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main/vae)
+with `--vae` to use the distilled video decoder. The ComfyUI FP16 and INT8
+ConvRot checkpoints include the MiniMax-H3 encoder for image and video inputs.
+Supply `--audio-vae` as usual to decode the soundtrack.
+
+Sol decodes the whole clip by default. Use `--vae-tiling` for spatial tiling or
+`--temporal-tiling` for temporal tiling. Either can reduce memory use but may
+affect output quality because the decoder has less context at tile boundaries.
+
+Temporal tiles default to 10 latent frames with 5 latent frames of overlap.
+Set `temporal_tile_frames` and `temporal_tile_overlap` in
+`--extra-tiling-args` to override them. Both must be multiples of 5 latent
+frames, and overlap must be smaller than the tile. The last tile may be shorter.
+Circular decoding requires spatial tiling on each circular axis, with a tile
+size smaller than the input on that axis.
+
 ## Latent hires second pass (two-stage sampling)
 
 MiniMax-H3 supports `--hires` with two interchangeable second-stage upscalers:

docs/performance.md

@@ -92,10 +92,12 @@ dimensions down to the minimum size. Each spatial retry must reduce the effectiv
 tile size. These runtime adjustments do not change the caller's parameters.
 Execution failures are not retried, and encoding has no automatic OOM retry.
 
-`--temporal-tiling` remains independent of spatial tiling. MiniMax H3 always uses
-spatial tiling (256x256 pixels and 25% overlap by default) and its own temporal
-windows. With `--vae-tiling`, its overlap follows `--vae-tile-overlap`; explicit
-spatial sizes are honored.
+`--temporal-tiling` remains independent of spatial tiling. The original MiniMax-H3
+video VAE uses spatial tiling (256x256 pixels and 25% overlap by default) and its own
+temporal windows. With `--vae-tiling`, its overlap follows `--vae-tile-overlap` and
+explicit spatial sizes are honored. [Sol-H3-VAE](minimax_h3.md#sol-h3-vae-decoding)
+decodes the whole clip by default and supports opt-in spatial and block-aligned
+temporal tiling. Tiling may affect output quality.
 
 ## Offload weights to the CPU to save VRAM without reducing generation speed.