Bulle d'Encre — a MOSS-Image preview
## Bulle d'encre is French for "ink speech balloon".
A MOSS-Image checkpoint with a style of its own: European comic ink and colour with an anime soul. It draws comic pages with panels, captions and speech balloons, album covers with real titles, and the clean, readable text MOSS-Image is known for. Tag prompts work too.
> This is a preview. Bulle d'Encre is an early look at the tests for a future HaDeS finetune of MOSS-Image. The full finetune, with a much larger dataset, is in progress. This file shows where it is heading, not the finished model.
## What it does well
Comic pages and covers: panel layouts, caption boxes, speech balloons with the exact words you write, album-cover titles.
Text: give it the exact words and it writes them, even small ones (every shirt in a locker room reading "LIONESSES").
A real European-comic look: ink lines, flat colour or watercolour, 1930s to 1960s adventure, noir and spy styles. It also does black-and-white pages.
Faces and ethnicities as described: very dark skin, elderly faces and named ethnicities follow the prompt.
Your Danbooru tags work: coming from Illustrious or NoobAI? Use the tags you know; you do not have to write paragraphs. Long natural-language prompts go further.
## Recommended settings
Steps: 15
CFG: 2.5 (up to 5 works too)
Sampler / Scheduler: Euler / sgm_uniform (simple also works)
Size: 1344 × 1792 (or any resolution higher than 2MP yields better images)
Flow shift: 5
Negative:
worst quality, bad anatomy, jpeg artifact
Text encoder: qwen3.5_9b_int8_convrot (Load CLIP, type moss_image). VAE: flux2-vae. Both from Comfy-Org/MOSS-Image.
No LoRA needed. Our Turbo v1.0 is already baked in at 0.5.
### Why the Turbo is baked in at half strength
At 0.5 the Turbo lets the model run at 15 steps and keep its style: it gets close to what the style gives at 25 steps without any Turbo. In our tests on the style LoRA, a low Turbo strength at 15 steps came closer to the 25-step image than no Turbo at all. At full strength the Turbo changes the style; different styles drift toward the same look.
Need speed? Load HaDeS MOSS Turbo 8-steps v1.0 at 0.5 on top (that makes the full Turbo) and use 4-8 steps. It works, with some loss of quality.
## The gallery
Every image uses the same seed, 1200, with the settings above. No seed hunting: these show what the prompt alone does. Each image says how it was prompted:
- Natural language + Prompt Enhancer: a long description written by the HaDeS MOSS Suite Prompt Enhancer, refined with a few correction notes.
- Tags: a plain Danbooru-style tag prompt.
The full prompt is in each image's metadata.
## How to prompt it
The gallery was made with the HaDeS MOSS Suite (free ComfyUI nodes, install it from ComfyUI manager or from oficial repo: https://github.com/Devildonia/HaDeS-MOSS-Suite): write the idea, let the Prompt Enhancer expand it, generate, then add notes after Corrections: and run it again, one fix at a time. Tips that worked:
Two characters: give each a place in the frame (*"in the lower foreground"*, "descending from above") and name the point of contact (*"his cutlass clashing against her rapier"*, "his hand grasps her wrist").
Text: write the exact words. One balloon per line, and say whose balloon it is and where.
Camera angles: write them as a full sentence (*"The perspective is an extreme low-angle worm's-eye view looking straight up…"*). A short tag like low angle is often ignored.
Describe what is there, not what is missing: "riderless" draws a rider. Metaphors are drawn literally.
## Known limits
Two characters interacting (holding, fighting, pulling) can need a correction or two, or another seed.
Brand logos can show up uninvited on everyday objects (coffee cups, a bus grille): describe the object plainly.
Now and then a fake signature, an extra copy of an object, or invented small text in the background.
It can produce mature content when asked.
## Files
BF16, 10.2 GB (the file the gallery was made with). It keeps its working name,
HaDeS_MOSS_preview_beta.safetensors, so the workflows in the gallery PNGs load it as they are.int8 ConvRot, 5.4 GB, for less VRAM. It is the BF16 file converted to Comfy-Org's int8 ConvRot layout, with no other change. Load it the same way, with the same settings (it needs a recent ComfyUI).
Which one? The int8 file is fully usable. We still recommend the BF16 if your GPU can hold it: it is the file the gallery was made with, and it gives exactly the gallery images from their metadata (re-rendered pixel for pixel).
What changes with int8, tested on six gallery prompts with the same seed and settings:
Same compositions in all six. The mean difference on a small thumbnail is 3.9 % (2.1-6.7 %).
Same text: the titles and balloons read the same ("BULLE D'ENCRE", "THE SECRET OF THE BLACK LIGHTHOUSE", the knight page, the duel).
Small details change: a pose shifts a little, the numbers on the shirts in a background differ.
Same sharpness.
So int8 images will be close to the gallery, but not identical to it.
Also on Hugging Face: DeVilDoNia/Bulle-d-Encre (full model card, recipe and measurements).
## Base Recipe
MOSS-Image-SFT with 30 % of MOSS-Image-Base in blocks 24-29, plus two HaDeS LoRAs baked in: Turbo 8-steps v1.0 at 0.5, and our style LoRA at 0.8, trained on 1200 images made with our own models. Saved with the HaDeS MOSS Suite Model Saver.
## Credits and license
Built on MOSS-Image-SFT and MOSS-Image-Base by OpenMOSS-Team, Apache 2.0. The LoRAs are our own. This checkpoint: Apache 2.0. Made with the HaDeS MOSS Suite for ComfyUI.
Feedback is welcome. The next version is being trained now. HaDeS.

