MOSS Pastel Bloom: a lighter, softer MOSS-Image, next to SFT and MossAD

Research note2026-11-02 · MOSS-Image Base (int8), MOSS-Image SFT (int8), MossAD-Demo (bf16), MOSS Pastel Bloom v1.0 (int8) · RTX 5070 12GB

One anime prompt, two seeds in columns, four MOSS-Image variants in rows: Base, SFT, MossAD with no trigger, and Pastel Bloom — Rooftop prompt, two seeds (columns), four variants (rows). Pastel Bloom is the bottom row.

MOSS Pastel Bloom is a third-party fine-tune of MOSS-Image SFT, published on Civitai by Aikimi on 2026-10-17 (model page). Its name says what it is for: soft pastel anime pictures. We ran it on the same prompts and seeds as our MOSS-Image anime comparison, and on our five standard prompts, so it now also has two columns in Compare.

What we ran

Same settings as SFT in the earlier note, with only the model file changed: 1024×1024, 25 steps, the er_sde sampler, CFG 2.5, the negative prompt from ComfyUI's MOSS-Image template, the same 4-bit Qwen3.5 9B text encoder and the same VAE. The model page does not recommend settings of its own, so we did not tune anything. We used the int8 file (5.3 GB); there is also a bf16 file (9.9 GB) that we did not try.

30 images: the five anime prompts with seeds 400 and 401, and both versions of our standard prompts with seeds 300 and 301. All on rented RTX 5070 cards with 12 GB of VRAM.

Variant	File size	Median time per image
SFT (int8), earlier run	5.4 GB	15.4 to 15.7 s
Pastel Bloom (int8), standard prompts v1	5.3 GB	23.4 s
Pastel Bloom (int8), standard prompts v2 and anime	5.3 GB	20.5 s

Pastel Bloom was slower than SFT, but these numbers come from three different machines of the same card type, and even the two Pastel Bloom runs differ by 14 %. The int8 files may also be packed differently. We did not find out why it is slower. The peak memory use was 11.2 to 11.7 GB of the 12 GB card.

The pictures

Each sheet is one prompt. Columns are two seeds, rows are the variants.

Rooftop prompt, four variants — Rooftop at sunset. Pastel Bloom keeps the red scarf and the city, but the sky and the buildings turn pale and the camera moves back.

Café prompt, four variants — Rainy café. Pastel Bloom's pictures are brighter and have less contrast than SFT's.

Elf archer prompt, four variants — Elf archer in a forest. Pastel Bloom draws the bow in both images, with a light green palette.

Festival prompt, four variants — Summer festival. SFT goes for close-ups under warm light; Pastel Bloom shows the whole scene with a cooler, blue night sky. Stall signs mix real-looking Japanese characters with distorted ones in every row.

Robot prompt, four variants — White and blue robot in a ruined city. Pastel Bloom keeps the same subject with paler smoke and sky.

What we noticed

The look changes the way the name says. Colours are lighter, contrast is lower and highlights go almost white. Faces are drawn with thinner lines and softer shading than SFT's, and the warm light SFT puts in the café and the festival turns cooler.
The camera moves back. In the rooftop and festival prompts, SFT often frames a close-up; Pastel Bloom more often shows the full figure and the background.
The prompt is still followed. Scarf, tea cup, bow, yukata and robot are in every Pastel Bloom picture.
It still makes photographs. On our standard prompts the fisherman, the landscape and the busy scene stay photographic, with a lighter, cooler colour balance. The "OPEN" sign is spelled correctly in both seeds, as with SFT.
No signatures that we saw. Unlike MossAD with an artist trigger, we did not see any signature or handle in these 30 pictures. We looked at the sheets by eye only, so small text could have been missed.

Standard prompts, SFT against Pastel Bloom — Our five standard prompts (v2, seed 300). Top row SFT, bottom row Pastel Bloom.

Limits: two seeds per prompt, the SFT settings rather than settings tuned for this model, and one card type.

Licenses and limits

MOSS-Image is Apache-2.0. Pastel Bloom's Civitai page allows commercial use of generated images and does not require credit; we name the author anyway. We have not checked what it was trained on and make no claim about that. We only ran the model and published pictures made from our own prompts. All 30 passed our automatic all-ages check (NudeNet) before publishing. The weights are not stored on this site.