Add basic Ministral 3 3B support (Ernie Image) - #436
Conversation
|
Co-authored-by: hpr <64584739+jarz76@users.noreply.github.com>
Do I need to download the mmproj file? |
|
@city96 Request to merge this into the main branch, thanks~ |
Look like there is no use for mmproj, Ernie image model family lacks editing support for now. In theory if "Ernie Image Edit" model comes out, maybe it's worth enabling mmproj loading now? But looking at qwen3vl loader logic answer is no. |
|
ops.py could be that it didnt work for me originally cause i run a rtx 2080 ti, but at least with this it works. |
I don’t think this is related to model quantization. Ernie’s image quality itself is quite poor — overly sharp and giving a dirty visual impression. You can reduce the steps from 8 to 4, and note that Ernie only shows better aesthetic performance at resolutions within 1K. At 2K, its aesthetics are also quite bad. |
Cherry-pick production fixes from city96/ComfyUI-GGUF open PRs: - city96#472 dequant device-constant cache (major LTX/sampling speedup) - city96#433 IQ1/IQ2/IQ3 torch dequant (extra TE quants) - city96#470 QK-norm .scale→.weight (silent NaN/black Flux-compat) - city96#467 dequant bare nn.Parameters (LTX learnable_registers) - city96#392 lumina2/zimage pad token shape fix - city96#456/city96#468 GGMLTensor dtype + dequantize() for core cast path - city96#461 WeightAdapter-aware move_patch_to_device - city96#469 force_patch on partial load/unload - city96#440/city96#436 mistral3 TE, city96#438 qwen35, qwen2 allowlist - city96#473 partial: Qwen3-VL deepstack mmproj map for MiniMax H3 TE Skipped mega/draft rewrites (city96#459, full city96#473 LazyGGUFReader, city96#336 Triton). See PR_BACKPORT.md for the full matrix.




Allows loading Ministral 3 3B model to inference Ernie Image and Ernie Image Turbo.
Used models:
Unet Model GGUF
CLIP Model Instruct GGUFCLIP Model Base GGUF Seems to work better that instruct
Default workflow provided in ComfyUI works just swap out Unet And Clip for GGUF loader nodes.