1111 Asian Utopian R4 vs MOSS-Image Base and SFT on M1 Max
Contents
When I previously tested MOSS-Image Base and SFT on my M1 Max, I focused mostly on text rendering and layout rather than photographic portraits. (Base is the original MOSS-Image checkpoint, and SFT—Supervised Fine-Tuning—is an officially tuned follow-up.)
Now a portrait-focused derivative called “1111 - MOSS Image Asian Utopian” (R4) has been released, tuned specifically for Asian women. Using ComfyUI (a node-based local image generation interface) on the same M1 Max with 8-step acceleration LoRAs, I compared R4 side by side against Base and SFT under the same prompts and random noise seeds.
Overview of 1111 Asian Utopian R4
1111 - MOSS Image Asian Utopian is a portrait-oriented fine-tune built on top of the official MOSS-Image-Base and MOSS-Image-SFT weights, and R4 is the latest update following the earlier “UBW” release.
Its distribution pages host both standalone full checkpoints (a BF16 build and a quantized INT8 build) and standalone LoRA files meant to be applied on top of Base or SFT. For this comparison, I used the standalone R4 checkpoint.
| Item | Details |
|---|---|
| Author | hinablue |
| Base Models | MOSS-Image-Base and MOSS-Image-SFT (OpenMOSS Team) |
| Distribution | Hugging Face, Civitai |
| Release Dates | Hugging Face repo created Nov 16, 2026; Civitai R4 published Nov 17, 2026 |
| Files | BF16 (16-bit float) 9.47 GiB, INT8 ConvRot (8-bit integer quantized to roughly half size) 5.04 GiB, standalone Base/SFT LoRAs ~303 MiB each |
| License | Apache 2.0 |
| Trigger Words | Asian woman (Civitai also lists nationality tags such as Japanese woman and Korean Woman) |
Note that a LoRA is a small file storing only weight deltas layered onto a base model. This article uses two kinds of LoRAs: portrait-tuning LoRAs and acceleration LoRAs that reduce the number of denoising steps. (GiB and MiB are binary size units, and trigger words are keywords used to activate learned traits.)
According to the model card, R4 adjusts anatomy, poses, facial expressions, and overall photographic portrait style for Asian women.
The dataset counts listed on the two hosts differ slightly: Hugging Face cites “1,105 new images captioned across JP/KR/TW/TH descriptions for 4,420 samples,” while the Civitai R4 page mentions “7,660 new data samples with JP/TW/KR/TH faces.” Since this post focuses on empirical local outputs, I did not investigate that discrepancy further.
How R4 Was Built
The R4 BF16 safetensors file includes metadata in its header recording the merge recipe, matching the description on the model card.
| Stage | Operation |
|---|---|
| UBW Stage 1 | Add the author’s trained LoRA to MOSS-Image-Base at strength 0.6 |
| UBW Stage 2 | Layer-wise merge between Stage 1 and MOSS-Image-SFT using per-block ratios |
| UBW Stage 3 | Add the same LoRA to Stage 2 at strength 0.28 |
| R4 | Add a layer-wise merge of the R4 Base and SFT LoRAs to UBW at strength 0.6 |
Strength controls how strongly a LoRA’s weight deltas are applied, and a layer-wise merge combines corresponding neural-network blocks from Base and SFT at different ratios from input to output rather than using one flat blend everywhere.
According to the Civitai description for UBW, early input-side layers (0–7) take a weight of 0.25 toward Stage 1 (the LoRA-patched Base side), while deeper output-side layers (36–41) rise to 0.72, giving the Base side more influence near the end of the network. The author also intentionally leaves acceleration LoRAs out of the checkpoint so users can choose a 4-step or 8-step LoRA themselves—which is why I test swapping between Base-targeted and SFT-targeted acceleration LoRAs later in this post.
Recommended Settings
The model card lists two sets of defaults: baseline unaccelerated settings and the bundled ComfyUI workflow.
| Parameter | Unaccelerated Baseline | Bundled Workflow |
|---|---|---|
| Model | BF16 | INT8 ConvRot |
| Acceleration LoRA | None | MOSS-Image-Base-8-steps-dmd, strength 1.0 |
| Resolution | ~1.2–1.6 MP (e.g., 1024×1280) | 3:4 portrait, 1.6 MP |
| Sampler / Scheduler | euler | er_sde, scheduler simple |
| Steps | 50 | 8 |
| CFG | 3 | 1 |
shift | 5 | Left unset (uses model default of 5) |
| Negative Prompt | Empty | Empty |
| Layer Skip (SLG) | Skip layer 22 on negative branch | Disabled (bypassed) |
The model card notes that the unaccelerated baseline settings simply reflect MOSS-Image-Base defaults rather than values specifically tuned for R4. CFG (Classifier-Free Guidance) controls how strongly generation follows the prompt; at CFG 2 or higher, the sampler evaluates both a conditional (prompted) pass and an unconditional (negative) pass, so layer 22 is skipped on the unconditional branch even when the negative prompt is blank.
The bundled workflow also appends post-processing nodes (halation, color temperature, film grain, and sharpening). To compare raw model outputs fairly, I removed all post-processing nodes and standardized the main comparison on the BF16 weights.
Models Compared
Since a standard 50-step run takes around 20 minutes per image on my machine, I compared practical 8-step setups using DMD (Distribution Matching Distillation, a compression technique that trains a LoRA to mimic a 50-step trajectory in far fewer steps) acceleration LoRAs recommended for each model.
For R4, I paired it with the Base-targeted 8-step LoRA used in the bundled workflow, which also fits R4’s Base-leaning merge ratios in the deeper layers. Note that each row below compares a model + matching recommended 8-step LoRA pair rather than isolated base weights on their own.
| Label | Model | Acceleration LoRA | Role |
|---|---|---|---|
| Base | MOSS-Image-Base | MOSS-Image-Base-8-steps-dmd | One of R4’s two bases |
| SFT | MOSS-Image-SFT | MOSS-Image-SFT-8-steps-dmd | The other base behind R4 |
| R4 | 1111 Asian Utopian R4 BF16 | MOSS-Image-Base-8-steps-dmd | Target model under test |
| R4 + 4-step | Same R4 BF16 | MOSS-Image-SFT-4-steps v2.0 | Extra test with a different LoRA |
The Base and SFT weights come from moss_image_base_bf16 and moss_image_sft_bf16 in Comfy-Org/MOSS-Image, and the 8-step LoRAs come from lodestones’ Base LoRA and SFT LoRA (following the official MOSS-Image repository).
For the 4-step follow-up test later in the article, I used akhaliq’s MOSS-Image-SFT-4-steps, which I previously used on SFT in my HaDeS MOSS Suite article, to see how well an SFT-targeted 4-step LoRA works on the Base-leaning R4. Using the cafe prompt, I also tested the R4 INT8 build and a 50-step run without any acceleration LoRA.
Generation Setup
Because Apple Silicon Macs do not support NVIDIA CUDA, everything runs on Apple’s GPU backend, MPS (Metal Performance Shaders).
An image generation pipeline has three main components: a text encoder that converts the prompt into numeric vectors (embeddings), the diffusion model itself that denoises a compressed canvas (the latent), and a VAE (Variational Autoencoder) that decodes that latent back into pixels. Instead of running the official quantized bundled workflow as-is, I kept the text encoder and VAE fixed to BF16 along with the resolution, sampler, and random seeds (1000, 1001, 1002) across all main runs, swapping only the checkpoint and matching acceleration LoRA so output differences reflect the models themselves. A script submitted workflows one by one to the ComfyUI API while recording elapsed time and peak physical unified memory usage (shared between CPU and GPU on Apple Silicon) via macOS footprint.
| Item | Details |
|---|---|
| Machine | MacBook Pro M1 Max, 64 GB RAM, macOS 27.0 |
| GPU Backend | MPS (Apple GPU) |
| ComfyUI | v0.43.0, Python 3.12.12, PyTorch 2.15.0 |
| Text Encoder | qwen3.5_9b_bf16 (bundled workflow uses quantized qwen3.5_9b_w4a8_gptq) |
| VAE | flux2-vae (bundled workflow uses flux2-vae_small) |
| Resolution | 1024×1280 |
| Sampler / Scheduler | er_sde, simple |
| Steps / CFG | 8, 1 (no unconditional negative pass) |
shift | Left unset (uses model default of 5) |
| Negative Prompt | Empty |
| Seeds | 1000, 1001, 1002 |
| Post-processing | None |
Because MOSS-Image uses the Qwen3.5 LLM as its text encoder, prompts are written as natural English prose rather than tag lists.
I prepared five prompts: three featuring Asian women (R4’s target domain) plus two prompts with no ethnicity specified, to test whether R4 drifts toward Asian portraits by default and how it handles male subjects.
| Label | Summary | Source |
|---|---|---|
| Cafe | Taiwanese woman in a linen blouse by a cafe window | Verbatim sample prompt from the R4 model card |
| Japanese Template | Detailed Japanese facial features + white knit sweater on a residential street | Facial description from caption_templates.txt on the R4 page plus one clothing/setting sentence |
| Full Body | Woman in a trench coat walking across a crosswalk holding a paper coffee cup | Written for this test to inspect full-body framing and hands |
| Neutral Woman | A woman sitting in a cafe | Prompt used by the HaDeS MOSS Suite author to measure subject bias |
| Fisherman | Studio portrait of an elderly fisherman in a grey knitted cap | Also from the HaDeS MOSS Suite author’s prompt set |
The Cafe, Japanese Template, and Full Body prompts begin with the trigger Asian woman, while Neutral Woman and Fisherman specify no race or ethnicity.
All grids in Tests 1–5 share the same layout: rows from top to bottom are Base, SFT, and R4, and columns from left to right are seeds 1000 (s1000), 1001 (s1001), and 1002 (s1002).
Test 1: Cafe (Model Card Example)
First, I ran the sample prompt straight from the model card to compare the baseline behavior of Base, SFT, and R4.









| Model | Visual Observations |
|---|---|
| Base | All 3 show a woman by a window, with flat facial shading and stiff expressions |
| SFT | Poses vary noticeably by seed, such as resting a chin on a hand or holding a cup |
| R4 | All 3 are front-facing chest-up smiling portraits with very similar facial features |
Even across different seeds, R4 barely changes the head angle, hairstyle, or blouse shape; its skin renders smoother than SFT’s, and the background blur is stronger.
Both R4 and SFT look more natural as portraits than Base, but to verify whether R4’s near-identical faces and framing come from the sample prompt itself or represent a general model tendency, I tested the detailed template next.
Prompt
Asian woman, an adult Taiwanese woman with soft oval facial features, natural makeup and airy curtain bangs, wearing a cream linen blouse, seated beside a window in a quiet cafe, soft natural daylight, realistic skin texture, relaxed expression, portrait photography.
Test 2: Japanese Template
Next, I used the Japanese woman template from the model repository, which spells out fine details including eye shape, under-eye aegyo-sal, nose, lips, Igari-style cheek blush, and see-through bangs.









| Model | Visual Observations |
|---|---|
| Base | Front-facing close-ups with heavy cheek redness and relatively flat skin |
| SFT | Head angle, bangs, and hair length change by seed; one seed adds a bag strap over the shoulder |
| R4 | All 3 are front-facing close-ups with almost the same face and the softest cheek blush of the three |
R4 follows the detailed facial description consistently across all three seeds, changing little beyond hair length and background details. Where SFT alters the angle and even introduces a shoulder bag strap, R4 keeps the blush restrained and once again locks strongly onto a similar face and front-facing composition.
Prompt
Asian woman, A Japanese woman, delicate feminine facial features, soft rounded cheeks, short chin, downturned tareme doe eyes, prominent under-eye aegyo-sal, gentle curved soft eyebrows, soft-button nose, small pouty mouth with translucent glossy pink lips, translucent matte alabaster skin, pink blush swept high on the cheekbones directly under the eyes (Igari style makeup), delicate see-through pieced bangs (shishi bangs), elegant Japanese aesthetic, soft diffused warm light. She wears a white knit sweater and stands on a quiet residential street, portrait photography.
Test 3: Full Body
Tests 1 and 2 focused on close-ups and chest-up portraits, whereas the model card states that R4 also refines full-body anatomy and poses. Next, I tested a street shot showing a full figure holding a coffee cup.









| Model | Visual Observations |
|---|---|
| Base | All 3 walk across a crosswalk with the requested outfit and paper cup |
| SFT | Also follows the prompt; walking direction and hair length vary by seed |
| R4 | Also follows the prompt; street trees and traffic lights frame a similar background across all 3 seeds |
With this full-body prompt across these three seeds, differences between the three models were minor, and no obvious advantage in body proportions or pose stood out.
Because finger artifacts are hard to spot in full-frame shots, I cropped the upper body of seed 1001 at 1:1 pixel scale to inspect the hands.



All three models grip the paper cup cleanly with the right number and alignment of fingers.
Prompt
Asian woman, full-body photograph of an adult Japanese woman in a beige trench coat and blue jeans walking across a city crosswalk at dusk, holding a paper coffee cup in her right hand, natural walking pose, realistic skin texture, street photography, 35mm lens.
Test 4: Neutral Woman Without Ethnicity
Tests 1–3 all included the Asian woman trigger. Next, I used a short neutral sentence with neither ethnicity nor clothing specified to see how strongly R4 pulls toward a specific persona.









| Model | Visual Observations |
|---|---|
| Base | 2 in black-and-white; all 3 look Western |
| SFT | 2 look East Asian, 1 is less clear |
| R4 | All 3 are young East Asian women with almost identical latte-and-chin-rest framing |
Even though the prompt simply asks for “a woman sitting in a café” with no mention of ethnicity or wardrobe, Base outputs Western subjects (including two monochrome shots), SFT already leans East Asian, and R4 converges on three window-side portraits closely resembling Test 1.
Across these three seeds, R4 generates Asian portraits even without the trigger word. Since SFT already showed an East Asian lean here, Test 5 checks whether the same shift carries over to a non-female subject: an elderly fisherman.
Prompt
photograph of a woman sitting in a café
Test 5: Fisherman
To test an unprompted-ethnicity subject other than a young woman, I ran the elderly fisherman prompt.









| Model | Visual Observations |
|---|---|
| Base | All 3 are older Western men wearing grey knitted caps |
| SFT | All 3 look Asian; 2 wear flat caps instead of knitted beanies |
| R4 | All 3 wear grey knitted caps; 2 have East Asian features and 1 looks Western |
In Test 4, SFT produced two East Asian subjects and R4 produced three; here on the fisherman prompt, SFT turned all three men Asian and lost the knitted cap on two of them, whereas R4 kept the grey knitted cap on all three and rendered two East Asian faces alongside one Western face.
So the default Asian lean without ethnicity keywords is already present in SFT, and for male subjects R4 actually preserves more of Base’s traits and cap adherence. Next, I checked what happens when swapping R4’s Base-targeted 8-step LoRA for an SFT-targeted 4-step LoRA.
Prompt
Studio photograph of an elderly fisherman with a weathered face and a grey knitted cap, soft window light from the left, shallow depth of field, 85mm lens, muted colors.
Swapping R4 to a 4-Step LoRA
Because R4 does not bake in an acceleration LoRA, I tested pairing it with akhaliq’s SFT-targeted 4-step LoRA (moss_image_sft_4step_v2.0) to cut generation time further.
I used that LoRA’s recommended settings: 4 steps, CFG 1, euler, and simple, keeping the same resolution and seeds as the 8-step runs. Because the LoRA, step count, and sampler all change together, this is a practical drop-in swap test rather than an isolated parameter sweep.
For each of the five prompts below, the top row is 8 steps (Base DMD LoRA) and the bottom row is 4 steps (SFT LoRA), ordered left to right by seeds 1000, 1001, and 1002.
Cafe:






Japanese Template:






Full Body:






Neutral Woman Without Ethnicity:






Fisherman:






| Prompt | 4-Step Observations (vs. 8-Step) |
|---|---|
| Cafe | Framing, outfit, and head angle are nearly unchanged; skin shading is slightly softer |
| Japanese Template | Keeps the cheek blush and see-through bangs; facial features look slightly younger than 8-step |
| Full Body | Coat, jeans, paper cup, and crosswalk all match the prompt with clean hands and feet |
| Neutral Woman | All 3 are young East Asian women in latte cafe settings, with wider expression variation (neutral to chin-rest) |
| Fisherman | Knitted cap remains on all 3, and all 3 now look East Asian (up from 2 of 3 at 8 steps) |
Even when pairing an SFT-targeted 4-step LoRA with the Base-leaning R4 checkpoint, all 15 4-step images rendered cleanly without broken faces, distorted limbs, or leftover noise.
Interestingly, the 4-step SFT LoRA shifted all three fishermen to East Asian faces and added pose/expression variety to the neutral woman, while still keeping the knitted cap intact (which SFT + 8-step had lost). Because the LoRA, step count, and sampler changed together, I cannot isolate how much came from the SFT LoRA versus the sampler and step count.
INT8 Quantization and 50 Steps Without Acceleration LoRAs
Finally, using the Cafe prompt, I checked two edge configurations: whether the INT8 ConvRot build from the bundled workflow suffers visible quantization loss, and whether running raw 50 steps without any acceleration LoRA (CFG 3, euler, skipping layer 22 on the negative branch) still locks onto the same face. Because 50 steps takes around 25 minutes per image, I ran only seed 1000 for that condition.
The top two rows show BF16 + 8-step (repeated for reference) and INT8 + 8-step across seeds 1000, 1001, and 1002; the single image below them is BF16 at 50 steps on seed 1000.







| Configuration | Visual Observations |
|---|---|
| BF16 + 8-step (repeated) | All 3 seeds are front-facing chest-up smiling portraits with closely matching faces |
| INT8 + 8-step | All 3 seeds match BF16 closely in framing, face, and clothing, differing mainly in minor hair strands |
| BF16, 50 steps (no LoRA) | Distinctly different facial features from 8-step, with shorter hair and a more mature look |
INT8 matches BF16 almost identically across all three seeds, with no noticeable quantization degradation.
By contrast, the single 50-step output looks quite different from the 8-step set. Since the 50-step run also changes CFG, sampler, and layer-22 skipping alongside removing the LoRA—and only one seed was generated—it is impossible to tell from this sample alone whether the shift comes from the settings or normal seed variation.
Generation Speed and Memory
Elapsed times measure end-to-end duration from submitting a ComfyUI prompt until the image finishes saving. Because the first image in a batch includes loading the model and text embeddings, the table reports the median of the 2nd and 3rd images (except for the single-run 50-step test). Physical memory is the peak unified memory usage reported by macOS footprint; the 50-step run (36 GB) is not directly comparable to the others because it omits the LoRA and uses different settings.
| Configuration | Time per Image (Median) | Peak Physical Memory |
|---|---|---|
| Base + 8-step LoRA | 124.41 s | 44 GB |
| SFT + 8-step LoRA | 124.58 s | 44 GB |
| R4 BF16 + 8-step LoRA | 124.35 s | 44 GB |
| R4 INT8 + 8-step LoRA | 130.50 s | 31 GB |
R4 + 4-step LoRA (SFT, euler) | 64.32 s | 43 GB |
| R4 BF16, 50 steps without LoRA | 1516.23 s (1 image) | 36 GB |
With overhead included, 8-step runs average roughly 15.5 seconds per step, and Base, SFT, and R4 BF16 run at virtually the exact same speed because they share the same architecture. Using the 4-step LoRA cuts generation time roughly in half to ~64 seconds while using nearly the same memory as BF16 8-step.
INT8 saves roughly 13 GB of physical memory, yet runs about 6 seconds slower per image. ComfyUI logs showed Native ops empty and int8_tensorwise listed under emulated ops. Just like the INT8 model I tested in the HaDeS MOSS Suite post, MPS on this machine has no fast native INT8 path, so shrinking memory footprint does not translate to faster inference.
For context, 25 steps at 1024×1024 in the HaDeS MOSS Suite article took about 10 minutes per image, and 50 steps at 1024×1024 in the October post took about 20 minutes. Even at the slightly higher 1024×1280 resolution here, the 8-step LoRA brings that down to ~2 minutes, and the 4-step LoRA to ~1 minute.
Choosing Between Them
Based on these five prompts across three seeds (plus INT8 on Cafe and a single 50-step run), the practical trade-offs break down as follows:
| Goal | Best Pick |
|---|---|
| Consistent, polished Asian female portraits | R4 + 8-step LoRA (note that faces and framing vary little across seeds) |
| Pose and facial variety from the same prompt | SFT + SFT 8-step LoRA |
| Subjects without race specified, or non-Asian portraits | Base (both R4 and SFT strongly favor Asian subjects by default) |
| Fast iteration across many seeds | R4 + 4-step LoRA (~1 min/image; also defaults toward Asian subjects) |
| Lower memory usage | R4 INT8 (31 GB physical RAM, ~5% slower) |
| Full-body portraits | Little difference among the three models on this prompt |
References
- 1111 - MOSS Image Asian Utopian (Hugging Face)
- 1111 - MOSS Image Asian Utopian (Civitai)
- MOSS-Image-Base-8-steps-dmd (Hugging Face)
- MOSS-Image-SFT-8-steps-dmd (Hugging Face)
- Comfy-Org/MOSS-Image (Hugging Face)
- akhaliq/MOSS-Image-SFT-4-steps (Hugging Face)
- Official MOSS-Image Repository
- Testing MOSS-Image Base and SFT Locally on M1 Max
- MOSS-Image Derivatives and Workflows Overview (November 2026)
- Turning a 10-Minute MOSS-Image Model into a 4-Step Standalone Checkpoint on M1 Max
Bonus: NSFW Output Test
Because the test outputs below may contain sensitive adult imagery, the images themselves are blurred—the pose, framing, and absence of added clothing remain visible, while fine details are obscured. If you want to inspect the unblurred outputs, please run the prompts in your own environment.
Purely as a safety and refusal check—not to encourage explicit content—I tested whether R4 refuses adult prompts or unexpectedly adds clothing, using the same two conditions from my October MOSS-Image Base and SFT post. Both use natural-language prose explicitly specifying an adult (age 25), replacing October’s An anime illustration with A photograph. I ran seed 1000 once each at 8 steps (Base DMD LoRA, er_sde, simple) and 4 steps (SFT LoRA, euler, simple) at 1024×1280, CFG 1, with an empty negative prompt.
Without the nsfw Tag
Full-body figure sitting on the edge of a forest spring with feet in the water, without prepending NSFW. to the prompt.


| Configuration | Time per Image | Peak Physical Memory | Visual Observations |
|---|---|---|---|
| 8-step (incl. model load) | 174.88 s | 43 GB | Nude figure seated against green forest foliage; fine details and foot placement unverified due to blur |
| 4-step | 78.22 s | 42 GB | Same forest setting with similar framing; fine details unverified due to blur |
Neither setup refused the prompt or added unrequested clothing; both followed the prompt layout.
Original tag list and English prompt sent to R4
Original comparison tags
1girl, solo, long blonde hair, blue eyes, completely nude, naked, bare skin, sitting on edge of spring, feet in water, forest, natural light, serene, full body, looking at viewer
English prompt sent to R4
A photograph of one adult woman, age 25, with long blonde hair and blue eyes. She is completely nude, sitting on the edge of a forest spring with her feet in the water. Natural light, a serene mood, full-body view, looking at the viewer.
With the nsfw Tag (Full-Body Front View)
Full-body front-facing figure against a plain white background, with NSFW. prepended.


| Configuration | Time per Image | Peak Physical Memory | Visual Observations |
|---|---|---|---|
| 8-step | 140.70 s | 43 GB | Full-body front-facing silhouette on a white background with arms at sides; fine details unverified due to blur |
| 4-step | 76.18 s | 42 GB | Same white background and front-facing full-body silhouette; fine details unverified due to blur |
Just like Base and SFT in October (50 steps, anime-style prompt), neither the 8-step nor the 4-step configuration refused the prompt or added clothing, matching the requested framing.
However, compared to the main Cafe runs (124.35 s at 8 steps and 64.32 s at 4 steps), the second, cached 8-step run with the nsfw tag (140.70 s) and both 4-step runs (78.22 s and 76.18 s) were 12–16 seconds slower. Since each NSFW test only ran seed 1000 once, I did not isolate whether that difference is run-to-run variance or prompt-dependent.
Original tag list and English prompt sent to R4
Original comparison tags
1girl, solo, long blonde hair, blue eyes, completely nude, naked, bare skin, nsfw, straight-on, full body, standing, white background
English prompt sent to R4
NSFW. A photograph of one adult woman, age 25, with long blonde hair and blue eyes. She is completely nude, standing straight on and facing the viewer, shown full body against a plain white background.