Testing MOSS-Image Base and SFT on an M1 Max Mac

I tested the public weights of MOSS-Image-Base and MOSS-Image-SFT locally on a 64GB M1 Max MacBook Pro using PyTorch’s MPS backend for the Apple GPU. OpenMOSS released the Base model and its supervised fine-tuned (SFT) variant on October 16, 2026.
Both models can draw text inside an image and accept experimental layout coordinates for placing objects. I tested the official examples alongside the character portraits and sunset scenes from my April Anima comparison and September Qwen-Image 2.1 evaluation. Base and SFT received identical inputs in each paired test; I show the earlier model images for reference.

MOSS-Image Base, SFT, and Recommended Settings

The official MOSS-Image repository provides download links for both Base and SFT weights. I evaluated character portraits, dynamic running scenes, in-image text rendering, and JSON layout conditioning on both models with identical prompts and parameters for each pair.

The Base weights and bundled pipeline were pinned to revision 5f40f07a304541840b6019d2d34d30f966756787, and the SFT weights to fc5de0d56fc910cbca6f6625a70f5f582b26deb1. The reference official GitHub repository was at commit 9d799c1e98a72d9f4e7c3d211072fd94cf365dee.
The file sizes across all 8 files matched the Hugging Face distribution manifest.

“5B” refers to the approximately 5 billion parameters in the image generation backbone. The text encoder converts the prompt into embeddings, the DiT progressively denoises the latent representation initialized from seed 42, and the VAE decodes the final latent into pixel space. In the downloaded weights, the text encoder took up more storage than the image generation backbone itself.

The table below summarizes the recommended settings from the official user guide. CFG controls how strongly the generation adheres to the prompt.
The model accepts natural language sentences, comma-separated tag lists, and JSON layout prompts. In JSON layout mode, positions for text and objects are specified with bounding box coordinates.

The step count determines how many times the noise is iteratively updated. Euler is a single-step numerical update method, and shift adjusts the timestep schedule distribution. I used the official defaults for both. With CFG, the model computes a conditional branch guided by the prompt and an unconditional branch (unconditioned if the negative prompt is empty, or conditioned on negative prompt tokens if provided). Skipping layer 22 was applied to the latter unconditional branch.

Comparison Baseline Models

The four baseline models use images previously published in the April and September articles, without re-generating them for this run. The comparison covers two main scenes: a full-body character portrait against a white background, and a dynamic figure running with a castle at sunset in the background. Both figures have blonde hair and wear white and gold robes. Because each model produced only a single image per condition, differences between individual samples cannot be generalized as an absolute ranking of model capability.

Generation Settings and Experimental Conditions

The MOSS character samples were generated at 832×1216, while the text rendering and poster samples were generated at 1024×1024. Both resolutions correspond to the officially recommended ~1-megapixel bracket. All MOSS runs shared 50 steps, CFG 3.0, shift 5.0, and seed 42. The seed is the starting point for the random values used in generation. I generated one sample per model for each condition. LoRA refers to auxiliary adapter weights trained for specific visual styles or subjects; no LoRA was used for MOSS in these tests. NSFW denotes adult content prompts.
The “UI initial prompt” in the table below means the quality and negative-tag strings prefilled in the official Gradio interface. I passed those strings to the Python API rather than generating through the interface.

The model cards for Base and SFT leave the default negative prompt empty in their Python code snippets. Conversely, the user guide recommends populating the negative prompt with quality-degrading terms, and the official Gradio UI pre-fills both positive quality tags and negative tags by default.
The MOSS guide notes that the model also accepts comma-separated tags. In my initial run, I translated the original tag list into natural English prose and explicitly added anime illustration. For those two initial Base/SFT samples, the negative prompt was left empty per the model card defaults, while uncond_skip_layers=[22] remained enabled. In subsequent tag-list tests, I incorporated the UI preset strings.
The UI preset strings went into the Python API arguments prompt and negative_prompt.

Across the seven conditions above, image resolution, 50 steps, CFG 3.0, shift 5.0, layer 22 skipping, and seed 42 were kept identical between Base and SFT. Only the two NSFW conditions were kept free from official UI quality and negative presets to observe how the model handled the prompt content directly.

Comparison images can be clicked to view full-size.

Base and SFT generate initial latent noise from seed 42 using the exact same pipeline. However, the four baseline models used different runtime implementations, step counts, and CFG values. Even with the same numeric seed, the initial noise in MOSS does not match the baseline models.

Furthermore, the official UI quality tags and negative prompt applied to MOSS cannot be retroactively applied to the previously generated Qwen and Anima outputs. The three April Anima-family models had no recorded negative prompt. Differences seen when comparing with baseline models therefore include variations in input conditions.

Test Environment and Launching on MPS

The test machine was a MacBook Pro with an Apple M1 Max and 64GB of unified memory running macOS 27.0. While the official documentation targets NVIDIA CUDA GPUs, I set up a dedicated Python 3.12.12 environment using Apple Silicon MPS (Metal Performance Shaders).
Execution was run with PYTORCH_ENABLE_MPS_FALLBACK=1 so that operations unsupported on MPS would fall back to the CPU.

The DiT and text encoder were loaded in bfloat16, while the VAE used float32. These settings represent values in 16-bit and 32-bit precision, respectively.

PyTorch handled GPU acceleration, and Diffusers invoked the official pipeline. Transformers was used to tokenize and encode text prompts. Swap refers to virtual memory written to disk storage when active memory demand exceeds physical RAM.

During the initial run, prompt processing halted because the manually downloaded files lacked the official tokenizer/chat_template.jinja. This file formats prompts for the text encoder. I added the Qwen3.5-9B template to chat_template in the local tokenizer_config.json. I later confirmed it was identical to the official Jinja template. Execution then halted on an unsupported float64 operation on MPS.

Timestep embedding tells the model which denoising stage the process is currently in. float64 and float32 are double-precision and single-precision floating-point formats, respectively. Because this error occurred even with fallback enabled, I applied the same patch to local working copies for both Base and SFT. The modification was strictly limited to routing that calculation through CPU; no model weights were altered. I did not generate images under identical conditions on CUDA, so I cannot determine whether MPS or the patch affected visual quality.
I compared the added template with the prompt formatting in the official ComfyUI node. For natural prose, tag lists, and JSON layout prompts, it produced the same <|im_start|>user format.
After transferring to the GPU, torch.mps.driver_allocated_memory() reported approximately 27.75GB allocated by the MPS driver.

Sending A cat at 512×512, 4 steps, CFG 3.0, shift 5.0, and seed 42 successfully saved a PNG file. Generation took 25.40 seconds, and model transfer to MPS took 9.00 seconds. MPS driver memory allocation was 27.75GB post-transfer and 30.14GB post-generation, with 0 swap usage measured after generation.
The 4-step image was blurry.

Official In-Image Text Rendering Prompt

I sent the sample prompt from the official user guide to both Base and SFT with the same line breaks. It asks for a refrigerator with a handwritten Chinese sticky note.
Settings were 1024×1024, 50 steps, CFG 3.0, shift 5.0, Euler, layer 22 skip, seed 42, and an empty negative prompt.
1024×1024 is within the guide’s recommended ~1-megapixel range and matches the resolution in the Python examples on the Base and SFT model cards.

Beneath each full image, I have included a 1:1 crop of the note magnified from the original full-resolution PNG, as the note text in the overview image is too small to read clearly.

MOSS-Image-Base
(User guide prompt)

Note crop

MOSS-Image-SFT
(Identical prompt)

Note crop

macOS OCR recognized the three requested Chinese lines from the full-resolution PNGs of both Base and SFT. Its confidence score is an estimate assigned to each line: the second Base line scored 0.50, while all three SFT lines scored 1.00.
A score of 1.00 does not guarantee that glyph shapes or punctuation marks match the prompt with complete fidelity. While the SFT note is readable across all three lines, it was not placed on the refrigerator door.
Generation time was 1,189.91 seconds for Base and 1,189.43 seconds for SFT, with 0 swap usage measured after both runs.

A wide view of a refrigerator beside a kitchen counter with a fruit bowl. A small sticky note on the refrigerator door is handwritten in Chinese:
“记得买牛奶和鸡蛋！
牛奶两盒，鸡蛋一打。
顺便带一袋面包。”

Specifying Poster Layout with JSON Bounding Boxes

I evaluated the JSON layout example from the official user guide on both Base and SFT.
The prompt specifies a 1024×1024 poster with a blue mountain illustration on a cream background and EXPLORE centered at the top. Bounding box coordinates are assigned separately to the text and mountain; the user guide notes this conditioning as experimental.
Identical JSON strings were sent to both models, and the negative prompt was left empty.

MOSS-Image-Base
(Official JSON)

MOSS-Image-SFT
(Identical JSON)

Measuring the text and mountain regions on the full-resolution PNGs using identical blue color threshold masks revealed that the deviation from the requested bounding boxes was at most 6 pixels for the text and at most 4 pixels for the mountain on both models. The guide’s bbox format is [top, left, bottom, right] with each axis spanning -512 to 512 around the center of a 1024×1024 image. I added 512 to each value to obtain the pixel coordinates measured from the image’s top-left corner in the table below. Because boundary values vary depending on how color thresholds are sampled, this is not a formal object detection metric.
SFT rendered a horizontal blue band across the bottom in addition to the mountain. Generation time was 1,228.94 seconds for Base and 1,236.49 seconds for SFT.

{"summary": {"meta": ["poster"], "canvas": "single", "caption": "A minimalist travel poster with a cream background, a blue mountain illustration, and the title EXPLORE at the top."}, "items": [{"label": "text", "caption": "Large dark-blue uppercase title centered at the top.", "text": "EXPLORE", "bbox": [-416, -384, -288, 384]}, {"label": "object", "caption": "A flat blue mountain illustration in the lower center.", "bbox": [-192, -288, 384, 288]}], "bbox_coord_space": "abs_centralized", "bbox_format": "yxyx"}

Comparing MOSS Character Portraits and Sunset Scenes with Four Baseline Models

In earlier comparisons, LoRAs were used to keep character identities and cafe backgrounds consistent. With no character LoRAs on hand for MOSS, these tests compared base foundation outputs directly.

Full-Body Standing Portrait on White Background

The initial concise English prompt with an empty negative prompt is shown later. For the primary comparison here, I prepended the official UI quality tags to the April tag list and used the five UI preset negative terms for both Base and SFT.

MOSS-Image-Base
(UI preset input)

MOSS-Image-SFT
(Identical input)

Qwen-Image 2.1
(September output)

Anima preview3-base
(April output)

WAI-Anima v1
(April output)

WAI-Illustrious v160
(April output)

Given identical tags, quality prefixes, and negative prompts, the SFT version distributed embroidery patterns all the way down to the hem compared to Base. However, both models showed predominantly only one shoe.
Compared to the previously published Qwen and Anima samples, both MOSS outputs had flatter shading across the face and white fabric. This is a reference comparison based on a single sample per model.

1girl, solo, long blonde hair, blue eyes, white robe, gold embroidery, capelet, gold sash, long sleeves, long dress, standing, looking at viewer, full body, white background

best quality, masterpiece, 1girl, solo, long blonde hair, blue eyes, white robe, gold embroidery, capelet, gold sash, long sleeves, long dress, standing, looking at viewer, full body, white background

Initial Test with Concise English Prose and Empty Negative Prompt

Prior to the primary comparison, I passed an identical concise English sentence to Base and SFT and left the negative prompt empty according to the model card defaults. Because the text encoder, tokenizer, VAE, and inference pipeline code were identical files in both distributions, they were shared across local working copies.

Expand initial Base and SFT outputs

Base

SFT

The English prompt sent to both models is below. The negative prompt was empty.

An anime illustration of one young adult woman, shown full body standing and looking at the viewer against a plain white background. She has long blonde hair and blue eyes. She wears a long white robe with gold embroidery, a capelet, a gold sash, long sleeves, and a long skirt.

In the initial test, Base showed sparser ornamentation on the shoulders and hem, whereas SFT added more intricate patterns to the shoulder capelet and waist sash. In both outputs, only one shoe was clearly visible. In the Base image generated with UI preset inputs, patterns further increased across the shoulders, chest, and front sash. Because prompt format, quality tags, and negative prompts were changed simultaneously, it is impossible to isolate which specific factor drove the increased patterning from comparing these two conditions alone. Generation times were 1,151.64 seconds for the initial Base run, 1,154.18 seconds for the initial SFT run, 1,159.20 seconds for Base with UI presets, and 1,156.02 seconds for SFT with UI presets.

Retrying SFT with Detailed English Prose

Using the same SFT model, I modified only the prompt. Following the detailed-description advice in the official Prompt Enhancer, I retained the core elements—blonde hair, blue eyes, white and gold robe, white background—while explicitly directing the model to render embroidery on the collar, capelet, sleeves, and hem, fabric folds and shadows, and both feet. Rather than running the skill’s JSON generation and compilation workflow, I wrote the English prose directly. Resolution at 832×1216, 50 steps, CFG 3.0, shift 5.0, layer 22 skip, seed 42, and an empty negative prompt were kept identical to the concise-prose SFT test.

While the concise-prose SFT sample clustered ornamentation primarily at the waist sash and capelet, this output extended gold filigree embroidery across the collar, capelet, sleeves, front panels, and hem, with both shoes clearly depicted. Shading across the face and fabric remains relatively subtle compared to the Qwen and Anima samples. I did not send the same detailed prompt to Base, so this difference cannot be attributed specifically to SFT. Generation time was 1,193.96 seconds (19 minutes 54 seconds).

A full-body anime character illustration of one young adult woman standing straight and facing the viewer against a seamless plain white background. Center the figure with her head and both shoes fully visible. She has long flowing blonde hair, blue eyes, and a calm expression. She wears an elaborate ankle-length white ceremonial robe made of layered ivory cloth, with a short shoulder capelet, wide long sleeves, a long skirt, and a patterned gold sash with a central clasp. Fine gold filigree embroidery follows the collar, capelet, cuffs, front panels, and hem. Show the individual ornamental motifs, layered fabric folds, and soft shadows that separate the white cloth from the white background. Clean, precise anime linework, restrained cel shading, warm highlights on the gold, even studio lighting. No props or scenery.

Figure Running with Castle and Sunset Background

I prepended the official UI quality tags to the April tag list and sent the same result to Base and SFT. The negative prompt used the five preset terms from the UI.

MOSS-Image-Base
(UI preset input)

MOSS-Image-SFT
(Identical input)

Qwen-Image 2.1
(September output)

Anima preview3-base
(April output)

WAI-Anima v1
(April output)

WAI-Illustrious v160
(April output)

Both Base and SFT captured the castle and sunset, but neither image clearly conveyed the arm swing or leg movement of a running stride. While SFT brought one shoe forward, both arms were outstretched sideways.
In contrast, the April WAI-Anima output showed a running posture with bent arms. MOSS generation times were 1,160.42 seconds for Base and 1,149.25 seconds for SFT.

1girl, solo, long blonde hair, blue eyes, white robe, gold embroidery, capelet, gold sash, long sleeves, long dress, running, wind, hair blowing, dynamic pose, fantasy landscape, castle in background, sunset sky, dramatic clouds, grass field

best quality, masterpiece, 1girl, solo, long blonde hair, blue eyes, white robe, gold embroidery, capelet, gold sash, long sleeves, long dress, running, wind, hair blowing, dynamic pose, fantasy landscape, castle in background, sunset sky, dramatic clouds, grass field

An anime illustration of one young adult woman with long blonde hair and blue eyes, running through a grassy field in a dynamic full-body pose. Wind blows her hair and the long sleeves of her white robe. Her robe has gold embroidery, a capelet, a gold sash, and a long skirt. A castle stands in the background beneath dramatic clouds and a sunset sky in a fantasy landscape.

Recreating Source Image Appearance Using Only Text Conditioning

In the April article, I evaluated image-to-image (i2i) generation at a denoise strength of 0.5 based on a character portrait generated with WAI-Illustrious v160 and the Kana-chan LoRA. Denoising strength controls how significantly the source image is transformed.
In the September article, I fed the same source image into Qwen-Image 2.1 to redraw it.

The source character has brown hair in a side ponytail tied with a blue scrunchie, amber eyes, a white and gold robe, and a large golden circle in the background.

Original WAI-Illustrious v160 source image

Before attempting the same experiment, I checked the call arguments of the official Diffusers pipeline. The pipeline’s __call__ does not accept an input image or a strength parameter; this call generates images from text.
Instead of providing the image, I described the hairstyle, blue scrunchie, costume, and background circle in English prose and generated a new image from text alone.

In this primary comparison, I prepended official UI quality tags to the descriptive English prompt and sent the same input to Base and SFT without passing the source image.
The earlier i2i outputs conditioned on the source image are displayed alongside for context, but this is not a direct comparison of identity retention.

MOSS-Image-Base
(Text only, UI preset input)

MOSS-Image-SFT
(Identical text)

Qwen-Image 2.1
(Fed source image)

Anima preview3-base
(denoise 0.5)

WAI-Anima v1
(denoise 0.5)

Both Base and SFT rendered the elements specified in the text: the brown side ponytail, blue hair tie, white and gold costume, and golden circle. SFT showed denser patterning on the shoulder capelet and hem.
The facial structure and clothing silhouette differ from the source image, but this does not measure i2i retention because no source image was passed to the pipeline. The prompt also did not specify the amber eye color of the original image.
Generation time was 1,147.47 seconds for Base and 1,143.18 seconds for SFT.

best quality, masterpiece, An anime illustration of a young adult woman with brown hair in a side ponytail tied with a blue scrunchie, standing full body facing the viewer. She wears a long white robe with gold embroidery, a capelet and a gold sash. A large muted gold circle is behind her against a pale background.

Generation Speed Benchmarks

I compared generation timing logged on the same M1 Max machine. The recorded seconds for MOSS measure the official pipeline’s generation call, which includes prompt encoding, iterative latent denoising, and VAE decoding. They exclude model loading, MPS device transfer, and PNG disk writes. Below are the measured timings across the seven conditions where Base and SFT received identical inputs.

According to the MOSS model card, the CFG 3.0 setting used here runs the image generation backbone twice per step. The earlier Qwen article used CFG 1.0.

The standing portraits, sunset scenes, text-described source tests, and NSFW runs were generated at 832×1216, while the refrigerator note and poster runs were at 1024×1024. Each row used 50 steps, CFG 3.0, and seed 42, with zero swap usage observed post-generation. Timings for the initial concise-prompt tests and the detailed-prompt SFT run are documented in their respective sections.

The speed comparison against earlier baseline models focuses on the white background standing portrait and dynamic running scene. MOSS and Qwen were run via Diffusers, whereas the other three models were executed in ComfyUI, with step counts and CFG values differing across models.

Model Choices for These Samples

These choices apply only to the single sample generated for each condition, not to an overall model ranking. The earlier models also used different prompts, LoRA settings, step counts, and CFG values.

Reference Links

• Official MOSS-Image repository (Release date, Base/SFT, Python examples)
• MOSS-Image-Base User Guide (Recommended settings and JSON layout)
• MOSS-Image-Base Model Card (Weights and license)
• MOSS-Image-SFT Model Card (Fine-tuned weights)
• Official Prompt Enhancer (Guidance for detailed descriptive prompts)
• Technical article comparing WAI-Anima with Anima and Illustrious (April portraits, running scenes, and i2i outputs)
• Technical article evaluating Qwen-Image 2.1 on M1 Max (September comparison images and generation timings)

Supplementary NSFW Output Tests

From this point forward, generated test images may contain sensitive material and have been blurred for publication. If you wish to examine unfiltered generation outputs, please test them in your own local environment.

I tested the same two content themes as in the April and September articles. For each theme, MOSS Base and SFT received the same English prose explicitly stating that the subject was an adult, age 25. Resolution, 50 steps, CFG 3.0, shift 5.0, and seed 42 matched the earlier character portrait runs.
Because the nsfw token in the official UI default negative prompt directly conflicts with the prompt content, the negative prompt was left empty for these two test runs. This is an explicit exception separate from the primary comparison using official UI quality presets. For the four baseline models, the previously published blurred images are shown.
The prompts also differ in pose and background, so these two images cannot isolate the effect of NSFW..

Without NSFW Tag

A full-body figure sitting at the edge of a forest spring with feet in the water, without passing the nsfw token.

MOSS-Image-Base
(This test)

MOSS-Image-SFT
(Identical input)

Qwen-Image 2.1
(September output)

Anima preview3-base
(April output)

WAI-Anima v1
(April output)

WAI-Illustrious v160
(April output)

Generation time was 1,132.28 seconds for Base and 1,133.58 seconds for SFT. The blurred images confirm the seated posture and forest background.

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

An anime illustration 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 NSFW Tag (Frontal Full Body)

A frontal full-body view against a plain white background, with NSFW. prepended to the English prompt.

MOSS-Image-Base
(This test)

MOSS-Image-SFT
(Identical input)

Qwen-Image 2.1
(September output)

Anima preview3-base
(April output)

WAI-Anima v1
(April output)

WAI-Illustrious v160
(April output)

Generation time was 1,131.51 seconds for Base and 1,132.87 seconds for SFT. While both outputs adhered closely to the requested composition, fine details of faces and artistic rendering remain filtered in the publication images.

1girl, solo, long blonde hair, blue eyes, completely nude, naked, bare skin, nsfw, straight-on, full body, standing, white background

NSFW. An anime illustration 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.