Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
136 commits
Select commit Hold shift + click to select a range
088032a
refactor(video): split H3BlockConfig out of H3Config
asher Nov 15, 2026
1e48eb9
refactor(video): take video sparse attention out of the H3 DiT block …
asher Nov 15, 2026
5e00e1a
refactor(video): move the H3 DiT's dtype, attention, linears, caption…
asher Nov 15, 2026
c5def73
refactor(video): int8 setting, refusals and pricing by title, variant…
asher Nov 15, 2026
ab5cf73
refactor(video): move the int8 setting, refusals and pricing to gmlx/…
asher Nov 15, 2026
cb5768c
refactor(video): move the H3 RoPE tables to gmlx/image/h3/rope.py
asher Nov 15, 2026
b6b4a4a
refactor(video): move the packed sequence layout to gmlx/image/h3/lay…
asher Nov 15, 2026
05ebd72
refactor(video): move the H3 block layer and H3Core to gmlx/image/h3/…
asher Nov 15, 2026
365eeac
refactor(video): move the H3 text encoder spec to gmlx/image/h3/encod…
asher Nov 15, 2026
b99afd7
feat(image): minimax-h3-image family row: fingerprint, config, pdd sa…
asher Nov 15, 2026
4265695
feat(image): minimax-h3-image dit: one latent frame on h3 blocks, pdd…
asher Nov 15, 2026
ab9fbbc
feat(image): minimax-h3-image loader by the family rule, check 3 on t…
asher Nov 15, 2026
e5900f3
refactor(video): H3 VAE token coordinates and RoPE tables on any numb…
asher Nov 15, 2026
db608fc
refactor(video): H3 VAE decoder blocks evaluated one at a time with a…
asher Nov 15, 2026
184d1cb
refactor(video): H3 VAE input normalization and tiled moments as func…
asher Nov 15, 2026
7ba255f
refactor(video): move the H3 VAE tile schedule and tile size to gmlx/…
asher Nov 15, 2026
25796cd
refactor(video): move the H3 VAE encoder layers, ViT blocks and tile …
asher Nov 15, 2026
946ae30
fix(image): minimax-h3-image cached layout positions read only
asher Nov 15, 2026
3ed1adc
test(image): minimax-h3-image check 4 on the pipeline's schedule and …
asher Nov 16, 2026
d9ecccd
fix(image): minimax-h3-image steps read the PDD heads of the grid int…
asher Nov 16, 2026
2c3517c
test(image): minimax-h3-image check 3 site gate on the release file, …
asher Nov 16, 2026
4a9308c
test(image): minimax-h3-image site gate faults print the GLU token ro…
asher Nov 16, 2026
2aa34ec
feat(image): MiniMax-H3-Image VAE row, always tiled at 256 px, with c…
asher Nov 16, 2026
0f7eb43
feat(image): reference encodes sample with the family's seed draws or…
asher Nov 16, 2026
c96ab11
feat(image): minimax-h3-image discovered, with its VAE and memory row…
asher Nov 16, 2026
33bf0d4
feat(image): size limits per image family, refused by name, with fitt…
asher Nov 16, 2026
3c3c21d
feat(image): sizes held to the family's limits in the CLI, the API sa…
asher Nov 16, 2026
5eb28f3
test(image): minimax-h3-image check 4 holds the head on steps that sp…
asher Nov 16, 2026
7c139d3
test(image): minimax-h3-image check 4 bound on the simple schedule re…
asher Nov 16, 2026
f2c8562
docs(image): drop a lab path from the minimax-h3-image VAE memory com…
asher Nov 16, 2026
feb68ae
refactor(image): minimax-h3-image check 3 loads the release file by d…
asher Nov 16, 2026
954e771
fix(image): minimax-h3-image decoder divides its four tiles a forward…
asher Nov 16, 2026
da58be0
fix(image): pack-vae and serve's part check read a VAE header under i…
asher Nov 16, 2026
1b9cc78
test(image): minimax-h3-image check 5 wrong models for the blend orde…
asher Nov 16, 2026
ea58a7f
fix(image): VAE GroupNorm statistics in float64 on the CPU, where sto…
asher Nov 16, 2026
9e92949
feat(image): minimax-h3-image step and encode prices refit from footp…
asher Nov 16, 2026
83b33a0
docs(image): minimax-h3-image size limits in the text under the reque…
asher Nov 16, 2026
d230e13
fix(serve): an image size that does not fit in memory names only a si…
asher Nov 16, 2026
befdfc3
docs(image): the step and VAE routes take a model's own size limits i…
asher Nov 16, 2026
a1721d7
docs(image): an edit or variation whose first image a model's own siz…
asher Nov 16, 2026
4aa5866
test(serve): the image routes test of a family's size limits runs its…
asher Nov 16, 2026
3131ffc
fix(image): bench asks for a size its family takes with --sizes, its …
asher Nov 16, 2026
fe4d681
docs(image): an unsized edit is scaled into its family's limits, as t…
asher Nov 16, 2026
c94b3f2
feat(image): a request with no size takes its variant's sampling row …
asher Nov 16, 2026
04e0894
feat(image): gmlx pull minimax-h3-image takes UD-Q4_K_XL on a 32 GB M…
asher Nov 16, 2026
af31059
feat(image): minimax-h3-image refuses steps other than 6 and guidance…
asher Nov 16, 2026
36170bf
fix(image): the minimax-h3-image VAE decodes a batch one image at a t…
asher Nov 16, 2026
f62a346
test(image): minimax-h3-image check 2, the Qwen3-VL-32B layer 50 tap …
asher Nov 16, 2026
c9e080d
fix(image): minimax-h3-image VAE prices refit on the largest of three…
asher Nov 16, 2026
17532fe
fix(image): the minimax-h3-image tiled decode blends each tile from t…
asher Nov 16, 2026
1a64a9c
test(image): minimax-h3-image check 6, the six PDD steps against Comf…
asher Nov 16, 2026
f6fe65b
fix(image): minimax-h3-image tiled decode prices refit on 17532fe9, w…
asher Nov 16, 2026
408bb9d
test(image): minimax-h3-image check 6 at 4096x4096, the exact run wit…
asher Nov 16, 2026
bccd4d8
fix(video): the minimax-h3 video tiled decode blends each tile from t…
asher Nov 16, 2026
385ef2d
fix(image): a pull prices the qwen3-vl-32b encoder it downloads as th…
asher Nov 16, 2026
0ae6692
test(image): the minimax-h3-image tiled decode blends each tile with …
asher Nov 16, 2026
7f6ddf2
test(image): the minimax-h3-image vae row covers every measured pass,…
asher Nov 16, 2026
7935882
test(image): the family prices fixture names the minimax-h3-image dec…
asher Nov 16, 2026
126b66a
docs(image): the minimax-h3-image license note cites each license sec…
asher Nov 16, 2026
e2b8732
fix(image): a request no size of whose shape fits names the true smal…
asher Nov 16, 2026
543c0ca
feat(image): minimax-h3-image runs its attention, its block linears o…
asher Nov 16, 2026
e16cb31
feat(image): gmlx image generate and bench take --int8 {off,attention…
asher Nov 16, 2026
4b71c56
feat(image): the image api and its steps take an int8 field that chan…
asher Nov 16, 2026
8366476
feat(image): the pack's model loader int8 setting reaches minimax-h3-…
asher Nov 16, 2026
304fbb3
test(image): minimax-h3-image check 3's teacher-forced block rows run…
asher Nov 16, 2026
fea7839
fix(image): a minimax-h3-image load at int8 linears frees each k-quan…
asher Nov 16, 2026
28ae8d2
fix(image): a minimax-h3-image load with its vae decodes in float16, …
asher Nov 16, 2026
db98524
docs(image): minimax-h3-image benchmark rows at 1024x1024 and 4096x40…
asher Nov 16, 2026
686707a
docs(image): the cli reference names --int8 on gmlx image generate an…
asher Nov 16, 2026
be66207
test(image): the family prices fixture carries the int8 step terms ea…
asher Nov 16, 2026
c89235d
fix(image): gmlx image bench --via-steps runs, prices and records the…
asher Nov 16, 2026
21d54dc
test(image): the minimax-h3-image tiled decode test pins the row-abov…
asher Nov 16, 2026
8fab675
fix(image): a minimax-h3-image size refusal and the no-fit 400 name t…
asher Nov 16, 2026
1361a6b
test(video): rewrap the docstring of the h3 decoder's one run canvas …
asher Nov 16, 2026
92960e9
feat(image): the image step route serves minimax-h3-image as comfyui'…
asher Nov 16, 2026
78e24ec
feat(image): the gmlx nodes run minimax-h3-image on comfyui 0.45 and …
asher Nov 16, 2026
71df1b0
test(image): the pack's minimax-h3-image int8 load allowance and the …
asher Nov 16, 2026
5abfadb
refactor(video): the H3 keyframe prompt code takes the label before e…
asher Nov 16, 2026
154ff70
refactor(image): move the H3 picture prompt code to gmlx.image.h3.pic…
asher Nov 16, 2026
e9e7729
refactor(image): move the H3 anchor rule, AnchorRule, augment and pin…
asher Nov 16, 2026
1dcbcd5
feat(image): the minimax-h3-image dit takes reference images as comfy…
asher Nov 16, 2026
a5b854c
test(image): check 3 holds a minimax-h3-image edit on the release fil…
asher Nov 16, 2026
d435aa9
fix(serve): a size that does not fit names the largest smaller size o…
asher Nov 16, 2026
d9b147b
feat(image): minimax-h3-image edits from up to 16 reference images by…
asher Nov 16, 2026
5335b40
feat(image): gmlx image generate --ref runs a minimax-h3-image edit a…
asher Nov 16, 2026
4d8f129
feat(serve): the edits and variations routes run minimax-h3-image edi…
asher Nov 16, 2026
52c0d0f
feat(serve): the step routes run minimax-h3-image edits as comfyui se…
asher Nov 16, 2026
733dbae
feat(comfyui): the pack runs minimax-h3-image edits through MiniMaxH3…
asher Nov 16, 2026
82366a2
fix(image): the vision tap drops its own references to each picture's…
asher Nov 16, 2026
d17724a
docs(image): the --ref row of docs/cli.md names minimax-h3-image and …
asher Nov 16, 2026
8dcdff3
fix(image): a minimax-h3-image prompt whose pictures have no referenc…
asher Nov 16, 2026
9ab8a00
test(image): check 6 on a minimax-h3-image edit at 1024, two pictures…
asher Nov 16, 2026
50d5cdb
docs(image): the --ref row of docs/cli.md fits the table's 160 charac…
asher Nov 16, 2026
981f5a8
test(image): the family prices fixture carries the minimax-h3-image r…
asher Nov 16, 2026
99b95b5
docs(comfyui): the pack lists seventeen example workflows
asher Nov 16, 2026
f8ffd20
fix(video): a bad keyframe's errors say keyframe again, the image's r…
asher Nov 16, 2026
c3b5729
feat(image): minimax-h3-image rewrites its prompt with MiniMax's expa…
asher Nov 16, 2026
772ab5b
feat(serve): a minimax-h3-image edit's rewrite sees its images throug…
asher Nov 16, 2026
82f3391
docs(image): docs/cli.md names --upsampler-mmproj and MiniMax's promp…
asher Nov 16, 2026
a50736b
docs(comfyui): the pack's parity section carries the minimax-h3-image…
asher Nov 16, 2026
c249cc6
feat(image): a minimax-h3-image rewrite closes its thinking at 3072 t…
asher Nov 17, 2026
adf3a25
docs(image): docs/images.md describes MiniMax-H3-Image's prompt expan…
asher Nov 17, 2026
99de25f
fix(serve): served image and video jobs run on the prefill command-bu…
asher Nov 17, 2026
820f62a
fix(serve): served image and video jobs run with the wired limit held…
asher Nov 17, 2026
72eebfd
docs(changelog): an image or video request served after a chat turn n…
asher Nov 17, 2026
4a52445
fix(serve): a served image or video request holds the wired limit at …
asher Nov 17, 2026
8dfabd0
fix(image): the CLI's expander turn reaches the copied mlx-vlm throug…
asher Nov 17, 2026
a7ef4a8
test(image): the family prices fixture carries the prompt_more field …
asher Nov 17, 2026
9ed74d4
refactor(video): the H3 LoRA map takes the audio modules and the audi…
asher Nov 17, 2026
da580ca
refactor(image): move the H3 LoRA key map and its refusals to gmlx.im…
asher Nov 17, 2026
8fcbdcf
feat(image): MiniMax-H3-Image takes LoRA files on H3's key map, a pai…
asher Nov 17, 2026
7cc3401
feat(image): the H3 LoRA map reads diffusers' MiniMax-H3 names, q, k …
asher Nov 17, 2026
064d4ae
fix(serve): the wired limit hold is one counted hold whose starts and…
asher Nov 17, 2026
7e52849
fix(serve): the step encode's rewrite runs at a fixed seed, so a Comf…
asher Nov 17, 2026
4e2c65e
test(image): check 7 holds minimax-h3-image's character-swap and pDMD…
asher Nov 17, 2026
a2b42f7
feat(comfyui): a minimax-h3-image LoRA in diffusers' names applies th…
asher Nov 17, 2026
9cfb5ac
fix(serve): an image request's int8 is no longer warned as an unsuppo…
asher Nov 17, 2026
a364229
feat(image): an adapter brings the schedule it was trained on from it…
asher Nov 17, 2026
a641726
feat(comfyui): gmlx PDD Heads has a MiniMax-H3-Image model read step …
asher Nov 17, 2026
b729e13
test(image): check 7 holds pDMD-2steps' two steps with step k reading…
asher Nov 17, 2026
8e1fb07
test(image): a LoRA in diffusers' names on MiniMax-H3-Image's int8 li…
asher Nov 17, 2026
e37eb64
fix(comfyui): the pack's LoRA key map finds its own wrap under anothe…
asher Nov 17, 2026
f68fc8a
docs: the adapter schedule, sigmas and int8 table rows fit the docs c…
asher Nov 17, 2026
86d59ea
test(image): check 7's delta half reads the original character-swap's…
asher Nov 17, 2026
ee2ad28
fix(image): a MiniMax-H3-Image LoRA in diffusers' names no longer war…
asher Nov 17, 2026
a90a8f8
fix(image): a LoRA spelling warning names stock ComfyUI, and says Com…
asher Nov 17, 2026
34268d0
fix(image): an adapter schedule takes each field from its entry, else…
asher Nov 17, 2026
7282d54
docs(image): MiniMax-H3-Image in docs/images.md, its text to image, 4…
asher Nov 17, 2026
16b6188
chore(merge): main ffcd8818 into h3-image
asher Nov 17, 2026
435d58c
test(image): the LoRA rules digest reads the spelling warnings that n…
asher Nov 17, 2026
cfaaa99
docs(images): the downloads of Z-Image-Turbo, MOSS-Image, FLUX.2 [kle…
asher Nov 17, 2026
b1f2449
fix(image): the MiniMax-H3-Image download counts its encoder's 1.1 GB…
asher Nov 17, 2026
f9371d1
fix(image): a LoRA spelling warning no longer doubles the period afte…
asher Nov 17, 2026
d4c9587
docs(images): the expander's GGUF comes with pull hf:, not pull minim…
asher Nov 17, 2026
6ed0115
docs(images): the first MiniMax-H3-Image command takes seed 3, whose …
asher Nov 17, 2026
22d22a1
chore(merge): main a06e2e4c into h3-image
asher Nov 17, 2026
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 19 additions & 0 deletions CHANGELOG.md
Original file line number Diff line number Diff line change
Expand Up @@ -32,6 +32,23 @@ adhere to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
images, video and sound on `/v1/systemone` and `gmlx systemone`, and
`gmlx pull` has them as `d1-2b` and `d1-omni-2b`. See
[Images, video and sound](docs/decisions.md#images-video-and-sound).
- `gmlx pull`, `image generate`, `serve` and the ComfyUI pack run
MiniMax-H3-Image on a Mac with 32 GB or more, at 2048x2048 or up to 20 MP,
with edits of up to 16 pictures. See
[MiniMax-H3-Image](docs/images.md#minimax-h3-image).
- MiniMax-H3-Image rewrites prompts with MiniMax's prompt expander, which sees
an edit's pictures. See [Prompt expander](docs/images.md#prompt-expander).
- `--int8 attention`, `linears` or `all` runs MiniMax-H3-Image in int8 on an M5
or newer. See
[Faster images with int8](docs/images.md#faster-images-with-int8).
- MiniMax-H3-Image LoRAs run in MiniMax's module names or diffusers', in gmlx
and through the gmlx ComfyUI nodes. See
[LoRA files](docs/images.md#lora-files).
- The ComfyUI pack has MiniMax-H3-Image workflows at 2048x2048 and 4096x4096,
an edit of two pictures, and pDMD-2steps in 2 steps with `gmlx PDD Heads`.
- An image LoRA such as pDMD-2steps runs on the schedule it was trained on,
and `--sigmas` or a request's `sigmas` set one. See
[Adapter schedules](docs/images.md#adapter-schedules).

### Changed

Expand Down Expand Up @@ -80,6 +97,8 @@ adhere to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).

### Fixed

- An image or video request served after a chat turn no longer takes tens of
GB more memory than the same request alone, nor wires all of it.
- Qwen3.8-Flash-Next requests that decoded together past 2052 tokens of
context attended to all of it and answered unlike a request decoding alone.
- Two Qwen3.8-Flash-Next requests failed when they decoded together with the
Expand Down
118 changes: 118 additions & 0 deletions docs/benchmarks.json
Original file line number Diff line number Diff line change
Expand Up @@ -3984,6 +3984,124 @@
"speedup_image": 1.93
}
]
},
"base-h3-image-1024": {
"model": "MiniMax-H3-Image pruned",
"steps": 6,
"size": 1024,
"rows": [
{
"engine": "ComfyUI with ComfyUI-GGUF",
"weights": "minimax_h3_image_pruned-UD-Q4_K_XL.gguf",
"step_s": 5.06,
"image_s": 31.07,
"gmlx_weights": "minimax_h3_image_pruned-UD-Q4_K_XL.gguf",
"gmlx_step_s": 1.15,
"gmlx_image_s": 7.29,
"speedup_step": 4.4,
"speedup_image": 4.26
},
{
"engine": "stable-diffusion.cpp",
"weights": "minimax_h3_image_pruned-UD-Q4_K_XL.gguf",
"step_s": 2.12,
"image_s": 13.97,
"gmlx_weights": "minimax_h3_image_pruned-UD-Q4_K_XL.gguf",
"gmlx_step_s": 1.191,
"gmlx_image_s": 7.54,
"speedup_step": 1.78,
"speedup_image": 1.85
}
],
"gmlx_int8": [
{
"int8": "off",
"step_s": 1.191,
"image_s": 7.54,
"speedup_step": 1.0,
"speedup_image": 1.0
},
{
"int8": "attention",
"step_s": 1.189,
"image_s": 7.52,
"speedup_step": 1.0,
"speedup_image": 1.0
},
{
"int8": "linears",
"step_s": 0.679,
"image_s": 4.51,
"speedup_step": 1.75,
"speedup_image": 1.67
},
{
"int8": "all",
"step_s": 0.684,
"image_s": 4.55,
"speedup_step": 1.74,
"speedup_image": 1.66
}
]
},
"base-h3-image-4096": {
"model": "MiniMax-H3-Image pruned",
"steps": 6,
"size": 4096,
"rows": [
{
"engine": "ComfyUI with ComfyUI-GGUF",
"weights": "minimax_h3_image_pruned-UD-Q4_K_XL.gguf",
"step_s": 39.213,
"image_s": 239.16,
"gmlx_weights": "minimax_h3_image_pruned-UD-Q4_K_XL.gguf",
"gmlx_step_s": 30.557,
"gmlx_image_s": 187.97,
"speedup_step": 1.28,
"speedup_image": 1.27
},
{
"engine": "stable-diffusion.cpp",
"weights": "minimax_h3_image_pruned-UD-Q4_K_XL.gguf",
"step_s": 91.045,
"image_s": 570.67,
"gmlx_weights": "minimax_h3_image_pruned-UD-Q4_K_XL.gguf",
"gmlx_step_s": 30.598,
"gmlx_image_s": 187.97,
"speedup_step": 2.98,
"speedup_image": 3.04
}
],
"gmlx_int8": [
{
"int8": "off",
"step_s": 30.598,
"image_s": 187.97,
"speedup_step": 1.0,
"speedup_image": 1.0
},
{
"int8": "attention",
"step_s": 28.393,
"image_s": 174.99,
"speedup_step": 1.08,
"speedup_image": 1.07
},
{
"int8": "linears",
"step_s": 24.705,
"image_s": 152.87,
"speedup_step": 1.24,
"speedup_image": 1.23
},
{
"int8": "all",
"step_s": 22.51,
"image_s": 139.75,
"speedup_step": 1.36,
"speedup_image": 1.35
}
]
}
}
}
Expand Down
70 changes: 63 additions & 7 deletions docs/benchmarks.md
Original file line number Diff line number Diff line change
Expand Up @@ -461,13 +461,13 @@ This comparison has its own settings in place of those under
## Image generation

gmlx draws a 1024x1024 image faster than ComfyUI and
stable-diffusion.cpp for Z-Image-Turbo and both Qwen-Image 2.1 models,
faster than ComfyUI for Krea 2 Turbo, and faster than mflux for
Z-Image-Turbo, on the M5 Max listed under
[Methodology](#methodology). Each speedup is the other engine's time
divided by gmlx's, with the same GGUF file on both sides, except for
mflux, which runs its own MLX quants against the gmlx file nearest in
bits.
stable-diffusion.cpp for Z-Image-Turbo, both Qwen-Image 2.1 models and
MiniMax-H3-Image, faster than ComfyUI for Krea 2 Turbo, and faster than
mflux for Z-Image-Turbo, on the M5 Max listed under
[Methodology](#methodology). MiniMax-H3-Image is also measured at
4096x4096. Each speedup is the other engine's time divided by gmlx's,
with the same GGUF file on both sides, except for mflux, which runs its
own MLX quants against the gmlx file nearest in bits.

- Each engine ran 1024x1024 images with seed 42, cfg 1, euler and the
model's default steps, one engine at a time in its own process, over
Expand Down Expand Up @@ -513,6 +513,62 @@ bits.
|---|---|--:|--:|--:|--:|--:|--:|
| ComfyUI with ComfyUI-GGUF | Krea-2-Turbo-Q4_K_M.gguf | 5.55 | 2.89 | 45.1 | 23.4 | 1.92x | 1.93x |

### MiniMax-H3-Image pruned, 6 steps

MiniMax-H3-Image ran at 1024x1024 and 4096x4096 on newer engine releases
than the models above, by the same rules, with these differences:

- ComfyUI 0.45.4 ran with ComfyUI-GGUF 373048b and torch 2.14.1 on MPS
in bf16, started with `--use-pytorch-cross-attention`, which took
less time a step than its default attention at both sizes, 17% less
at 4096x4096. It read a copy of the GGUF with one more metadata key,
the six-step schedule that ComfyUI reads from the file, with every
tensor the same. ComfyUI-GGUF loads no Qwen3-VL-32B encoder GGUF, so
ComfyUI took the prompt conditioning from gmlx, and its times cover
the DiT, the sampler and the VAE.
- stable-diffusion.cpp b9d7176b was built with Metal and run with
`--diffusion-fa` and `--vae-tiling`, with the
qwen3vl_32b_minimax_h3 Q4_K_M encoder its guide names. gmlx ran the
Qwen3-VL-32B-Instruct Q4_K_M encoder. The seconds per image of
stable-diffusion.cpp leave out its prompt encode, which it runs once
for a batch.
- At 1024x1024 each round timed two images after one or two untimed
ones. At 4096x4096 it timed one image after one untimed image,
ComfyUI's at 1024x1024.
- gmlx ran beside ComfyUI in one set of rounds and beside
stable-diffusion.cpp and its own `--int8` settings in another, so
each row gives the gmlx figures of its own rounds.
- The second table at each size gives gmlx's `--int8` settings on the
same file, each one's speedup taken over `--int8 off`.

At 1024x1024:

| Engine | Weights | Seconds per step | gmlx | Seconds per image | gmlx | Speedup per step | Speedup per image |
|---|---|--:|--:|--:|--:|--:|--:|
| ComfyUI with ComfyUI-GGUF | minimax_h3_image_pruned-UD-Q4_K_XL.gguf | 5.06 | 1.15 | 31.1 | 7.3 | 4.40x | 4.26x |
| stable-diffusion.cpp | minimax_h3_image_pruned-UD-Q4_K_XL.gguf | 2.12 | 1.19 | 14.0 | 7.5 | 1.78x | 1.85x |

| gmlx setting | Seconds per step | Seconds per image | Speedup per step | Speedup per image |
|---|--:|--:|--:|--:|
| `--int8 off` | 1.19 | 7.5 | 1.00x | 1.00x |
| `--int8 attention` | 1.19 | 7.5 | 1.00x | 1.00x |
| `--int8 linears` | 0.68 | 4.5 | 1.75x | 1.67x |
| `--int8 all` | 0.68 | 4.6 | 1.74x | 1.66x |

At 4096x4096:

| Engine | Weights | Seconds per step | gmlx | Seconds per image | gmlx | Speedup per step | Speedup per image |
|---|---|--:|--:|--:|--:|--:|--:|
| ComfyUI with ComfyUI-GGUF | minimax_h3_image_pruned-UD-Q4_K_XL.gguf | 39.21 | 30.56 | 239.2 | 188.0 | 1.28x | 1.27x |
| stable-diffusion.cpp | minimax_h3_image_pruned-UD-Q4_K_XL.gguf | 91.05 | 30.60 | 570.7 | 188.0 | 2.98x | 3.04x |

| gmlx setting | Seconds per step | Seconds per image | Speedup per step | Speedup per image |
|---|--:|--:|--:|--:|
| `--int8 off` | 30.60 | 188.0 | 1.00x | 1.00x |
| `--int8 attention` | 28.39 | 175.0 | 1.08x | 1.07x |
| `--int8 linears` | 24.71 | 152.9 | 1.24x | 1.23x |
| `--int8 all` | 22.51 | 139.8 | 1.36x | 1.35x |

## Serving measurements

These measurements back the performance pages' guidance. Unless a
Expand Down
Loading
Loading