Skip to content

Add basic Ministral 3 3B support (Ernie Image)#436

Open
deinferno wants to merge 2 commits intocity96:mainfrom
deinferno:main
Open

Add basic Ministral 3 3B support (Ernie Image)#436
deinferno wants to merge 2 commits intocity96:mainfrom
deinferno:main

Conversation

@deinferno
Copy link
Copy Markdown

@deinferno deinferno commented Apr 14, 2026

Allows loading Ministral 3 3B model to inference Ernie Image and Ernie Image Turbo.

Used models:
Unet Model GGUF
CLIP Model Instruct GGUF
CLIP 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.

Comment thread loader.py Outdated
@jarz76
Copy link
Copy Markdown
Contributor

jarz76 commented Apr 15, 2026

Reference image:
image
Prompt

一张排版精美的发型历史图鉴插画,采用 2 行 5 列的横向网格布局,展示了同一个女性在不同历史时期的 10 种经典发型。背景为柔和的浅米色,网格边框为极细的黑线。画面中的女性面貌特征高度统一,拥有浅色肌肤、深邃的眼睛和柔和自然的妆容,皆身穿简约的白色无袖上衣以凸显头部造型。每个格子内上方为女性头部与肩部的特写画像,下方为对应的发型名称和简短历史介绍文字,排版居中且整齐。

第一行从左至右依次为:
1. 第一格:女性梳着维多利亚时代的复古盘发,头顶有复杂的编发和盘卷。下方加粗标题文字为 'Victorian Updo',正文说明文字为 'Popularized in the 19th century, characterized by intricate braids and coiled buns, reflecting the era's modesty and elegance.'
2. 第二格:女性留着经典的 20 年代波浪短发,深色发丝紧贴头皮呈现优美的 S 型波纹。下方标题为 '1920s Finger Waves',正文说明为 'A hallmark of the Flapper era, created by pinching the hair into s-shaped waves close to the head.'
3. 第三格:女性头顶两侧各有一个巨大的向内卷曲的圆筒状发髻,为 40 年代胜利卷风格。下方标题为 '1940s Victory Rolls',正文说明为 'Iconic WWII era style featuring voluminous upward curls, symbolizing patriotism and practical beauty.'
4. 第四格:女性留着极短的精灵头,带有短短的锯齿状刘海。下方标题为 '1950s Pixie Cut',正文说明为 'A rebellious yet chic short crop made famous by Hollywood actresses, breaking away from traditional long hair.'
5. 第五格:女性头顶高耸着巨大的蜂窝状盘发,发量极显蓬松。下方标题为 '1960s Beehive',正文说明为 'A towering, backcombed style developed in 1960, symbolizing the futuristic and bold fashion of the decade.'

第二行从左至右依次为:
6. 第六格:女性留着向外翻卷的金色层次长发,呈现 70 年代羽毛剪风格。下方标题为 '1970s Feathered Hair',正文说明为 'Defining the disco era, this bouncy, layered look was heavily inspired by pop culture icons of the time.'
7. 第七格:女性留着极其蓬松、体积庞大的爆炸卷发,体现 80 年代狂野风格。下方标题为 '1980s Big Hair',正文说明为 'Characterized by excessive volume and hairspray, reflecting the bold, loud, and expressive culture of the 80s.'
8. 第八格:女性留着及肩的中长直发,发尾向内微扣,修剪得层次分明。下方标题为 '1990s The Rachel',正文说明为 'A bouncy, square-layered haircut defining 90s television fashion and beloved by millions of women.'
9. 第九格:女性梳着光洁的高马尾,额头两侧垂下两缕细长的发丝。下方标题为 '2000s Y2K Ponytail',正文说明为 'A sleek updo with face-framing tendrils and butterfly clips, capturing the playful early internet aesthetic.'
10. 第十格:女性留着现代法式慵懒风的棕色中长微卷发,自然随性。下方标题为 'Modern Beach Waves',正文说明为 'A contemporary favorite focused on effortless, messy texture, highlighting natural beauty and low maintenance.'

整体画风细腻写实,光线均匀柔和,精准照亮人物面部和每一丝发型的质感纹理,具有时尚杂志般的复古图鉴排版质感。所有文字字迹清晰可见,标题字体加粗醒目,正文采用易读的黑色无衬线字体。

Safetensor
python_2oHOFHrRy0

w/o llama permute
python_A4Xxda4RPW

w llama permute applied
python_HdRHRaTW1z

Co-authored-by: hpr <64584739+jarz76@users.noreply.github.com>
@Amazon90
Copy link
Copy Markdown

Allows loading Ministral 3 3B model to inference Ernie Image and Ernie Image Turbo.

Used models: Unet Model GGUF CLIP Model GGUF

Default workflow provided in ComfyUI works just swap out Unet And Clip for GGUF loader nodes.

Do I need to download the mmproj file?

@Amazon90
Copy link
Copy Markdown

@city96 Request to merge this into the main branch, thanks~

@deinferno
Copy link
Copy Markdown
Author

Do I need to download the mmproj file?

Look like there is no use for mmproj, Ernie image model family lacks editing support for now.
Loading mmproj seems to work if forced by editing code.

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.

@ThuGie
Copy link
Copy Markdown

ThuGie commented Apr 15, 2026

ops.py
ernie.json
with this loading ernie gguf should work, worked for me. but quant quality is bad, though cant compare to full version.

could be that it didnt work for me originally cause i run a rtx 2080 ti, but at least with this it works.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants