MiniMax-H3-Image for ncnn

BF16 ncnn weights for MiniMax-H3-Image text-to-image and reference-image editing.

Split DiT deployment (2026-11-16)

The prepared transformer core and shared BF16 conditioning banks for 4, 5 and 6 steps are available under h3image/transformer/:

File Size Purpose
transformer-core.ncnn.bin + .param 40.14 GB Image/text-dependent DiT core
fixed-modulations-bf16.bin 349 MB Default released six-step schedule
fixed-modulations-bf16-steps4.bin 233 MB Experimental four-step schedule
fixed-modulations-bf16-steps5.bin 291 MB Experimental five-step schedule
fixed-build/modulations.ncnn.bin + .param 26.11 GB Producer for online conditioning at any step count

The core copies retained checkpoint weights byte for byte. These banks preserve the tensor payloads from the validated presets; only their metadata was updated to make model identities independent of download timestamps. CPU and Vulkan use the same BF16 bank. Different execution backends can still produce different final images. Four and five steps change the released six-step schedule and remain experimental.

Use an updated h3image runtime supporting the unified split DiT layout, portable H3-fixed-model-v2 identities and H3FIXED bank formats 3/4. The default six-step bank does not require online conditioning. Both conditioning modes use the same core; neither requires an additional full transformer copy.

Download only the prepared deployment

This downloads approximately 93.46 GB (87.04 GiB), including the other required model components and all three banks. It excludes the original full transformer and the optional conditioning producer:

hf download nihui-szyl/h3-image-ncnn --local-dir models \
  --include "h3image/processor/*" "h3image/text_encoder/*" \
  "h3image/vision/*" "h3image/vae/*" \
  "h3image/transformer/transformer-core.ncnn.*" \
  "h3image/transformer/fixed-modulations-bf16*.bin"

Select models/h3image with the runtime's -m option. Keep the core checksum sidecar and the banks together. Verify new DiT file checksums against dit-SHA256SUMS; dit-manifest.json records sizes, identities, schedules and unchanged tensor-record checksums. The runtime trusts verified checksum sidecars instead of rehashing large weights at every start. Keep model files immutable during use.

Online conditioning

For online conditioning, additionally download the split producer (about 26.11 GB):

hf download nihui-szyl/h3-image-ncnn --local-dir models \
  --include "h3image/transformer/fixed-build/*"
  • --fixed-modulations 1 (default) uses a compatible precomputed bank. If no compatible bank is available, it computes conditioning online with the producer, including for six steps.
  • --fixed-modulations 0 ignores precomputed banks and computes conditioning online for the selected step count.

Both modes support text-to-image and reference-image editing. Online conditioning runs before sampling, follows the selected RAM/VRAM budgets, and releases its producer weights before loading the core. Its temporary BF16 bank is removed when the model instance is destroyed. Recomputing on a different backend can round conditioning values differently; keep the same validated banks when reproducing a prepared preset.

The 40.14 GB core and 26.11 GB producer contain the original DiT tensor bytes split between two graphs. The redundant transformer.ncnn.bin, transformer.ncnn.param and transformer.ncnn.bin.sha256 have been removed from this model repository. Keep transformer-core.ncnn.bin.sha256, the producer checksum and fixed-build/source.sha256 provenance sidecar. Core, producer and existing 4/5/6-step bank payloads are unchanged.

The six-step core and compatible bank support both text-to-image and reference-image conditioning. They do not make the 40 GB core fit entirely in a small GPU: the native runtime streams weights when necessary, and speed depends on available RAM/VRAM and storage.

Downloads last month
38
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support