Tech10 min read

Speeding Up MOSS-Image to 50s on M1 Max and Building a Standalone Model with HaDeS MOSS Suite

IkesanContents

Update (2026-11-18): Applied the SFT 4-step LoRA, INT8 MPS comparison, and HaDeS bias-check prompts from this post to benchmark the “1111 Asian Utopian R4” derivative against Base and SFT on M1 Max → 1111 Asian Utopian R4 vs MOSS-Image Base and SFT on M1 Max

When I tested MOSS-Image Base and SFT on an M1 Max, generating a single 1024×1024 image at 25 steps took nearly 10 minutes.
MOSS-Image is an open-weights diffusion model with impressive typography and photorealistic rendering capabilities, but running the standard 25 sampling passes on Apple Silicon makes iterative generation impractical.

The primary speedup comes from a 4-step distillation LoRA, which allows the diffusion process to converge in just 4 steps.
On November 3, 2026, HaDeS MOSS Suite was released as a custom node pack for ComfyUI. Rather than generating images directly, it provides tools for inspecting, pruning, and baking MOSS-Image weights and LoRA adapters. The suite includes:

NodeFunctionality
MOSS Model Block Merger, Multi-Model MergerMerges models using block-level ratios or techniques like TIES
MOSS LoRA ExtractorExtracts differences from fine-tuned weights as compact LoRA adapters
MOSS LoRA Block WeightApplies LoRAs with custom scaling across 6-layer blocks
MOSS LoRA ResizeShrinks LoRA rank to reduce file size
MOSS LoRA MergerCombines two LoRA adapters into one
MOSS Model SaverBakes LoRA weights into the base diffusion model and saves as a single safetensors file

The author evaluated these nodes on an RTX 5090 and published the benchmarks in tools/RESULTS.md.
Among the findings was an analysis of the community 4-step distillation LoRA: out of 42 transformer layers, the adapter acts primarily on layers 24–41. Compressing those layers down to rank 2 preserves the visual quality of the full 4-step LoRA.

Applying this on a Mac cuts generation time from 10 minutes to roughly 50 seconds while dramatically reducing adapter overhead.
I replicated the author’s benchmark conditions on an M1 Max, pruned the LoRA to rank 2, and merged it into an int8 standalone diffusion model file.

Test Procedure

The adapter used is akhaliq’s MOSS-Image-SFT-4-steps (moss_image_sft_4step_v2.0.safetensors, 1.19GB, rank 128, Apache 2.0).
It distills the MOSS-Image SFT model to converge in 4 sampling steps without output degradation.

Following the author’s benchmark workflow, the evaluation proceeded in four stages:

StageNodes UsedVerification Goal
1Standard LoRA LoaderVerify whether 4-step LoRA accelerates generation without artifacts
2LoRA Block WeightConfirm whether LoRA activity is concentrated in layers 24–41
3LoRA ResizeCheck if reducing rank to 8 or 2 maintains identical imagery and text rendering
4LoRA Block Weight, Model SaverBake the pruned LoRA into SFT, save as int8, and generate with standard nodes

Verification Environment

ItemDetails
HardwareMacBook Pro M1 Max, 64GB Unified Memory, macOS 27.0
GPU BackendMPS (Apple Silicon GPU)
ComfyUIv0.43.0 (MOSS-Image supported from v0.42.0), Python 3.12.12, PyTorch 2.15.0
Custom NodesHaDeS MOSS Suite 88839e0 (as of November 5, 2026)
Diffusion Modelmoss_image_sft_bf16 (Comfy-Org/MOSS-Image, ~10.2GB)
Text Encoderqwen3.5_9b_bf16 (~19.3GB)
VAEflux2-vae
Baseline Sampling1024×1024, er_sde, simple, 25 steps, CFG 2.5, shift 5.0
4-step Sampling1024×1024, euler, simple, 4 steps, CFG 1 (matching author settings)

Because my primary ComfyUI installation preceded MOSS-Image support, I set up a clean v0.43.0 environment.

For generation and metric calculation, I used the author’s evaluation script tools/moss_eval.py.
The script sends graph workflows directly to the ComfyUI API, allowing identical prompts, seeds, and image similarity metrics to run locally. I only needed to update the LoRA file path.
Additional tasks not covered by moss_eval.py—such as LoRA Resize and Model Saver serialization—were executed via helper scripts importing the original evaluation functions.

Prompts matched the author’s suite: two photographic portraits (“woman at a cafe”, “fisherman portrait”), one watercolor illustration (“fox under mushroom”), and three typography tests (“bakery wooden sign”, “three-line poster”, “layout JSON poster”).
Seeds 1000–1002 produced 9 photo/illustration samples and 9 typography samples per configuration.

Two environment differences exist compared to the author’s RTX 5090 setup:
The author used an int4 text encoder, whereas I ran bf16.
The author used SageAttention, while I relied on PyTorch’s native SDPA attention since SageAttention does not support MPS.

Image variance was measured using the author’s four metrics: Composition (luminance difference downscaled to 32×32), Edge (Sobel edge cross-correlation), Texture (high-frequency spectrum difference), and Color (mean and variance difference in Lab color space). Zero indicates an identical match.
As a baseline reference, changing only the seed on full-layer LoRA generations yielded an average composition delta of 16.49.

1. Speedup with the 4-Step LoRA

I generated 9 samples across three baseline conditions: 25-step standard SFT, 4-step LoRA, and 4-step generation without LoRA.

25 steps (Baseline)
Watercolor fox reading a book under a mushroom, generated at 25 steps
4-step LoRA
Watercolor fox generated with 4-step LoRA
4-step Without LoRA
Blurry corrupted image from 4 steps without LoRA
25 steps (Baseline)
Photograph of a woman at a cafe, generated at 25 steps
4-step LoRA
Photograph of a woman at a cafe, generated with 4-step LoRA
4-step Without LoRA
Unidentifiable brown image from 4 steps without LoRA

Without the LoRA, 4-step generation completely broke down into blurry noise across all 9 images.
With the 4-step LoRA applied, all 9 outputs rendered cleanly with proper composition and fine texture.
Because the sampler and step count differ from the 25-step baseline, outputs with the same seed yield different compositions.

ConfigurationTime per Image (Median)
SFT, 25 steps (Baseline)574.65s
SFT + 4-step LoRA46.93s
SFT, 4-step without LoRA46.97s

Times represent the median across 6 warmed-up runs (model and text already cached), including sampling and VAE decode.
The 4-step LoRA accelerated per-image generation from 574.65 seconds to 46.93 seconds—a ~12x speedup.
On the author’s RTX 5090, latency dropped from 15.14s to 1.54s (~9.8x), showing consistent relative scaling.

Beyond reducing steps from 25 to 4, setting CFG to 1 skips negative prompt passes, roughly halving per-step latency from ~23s to ~12s on M1 Max.

In the cafe portrait at seed 1001, a semi-transparent ghost of typography overlapped the foreground.

4-step LoRA, Seed 1001
Double-exposure ghosting artifact over portrait

The author reported an identical double-exposure ghosting artifact with this LoRA, confirming consistent behavior on Apple Silicon.

Physical memory consumption was monitored via macOS footprint every 20 seconds.
Peak memory reached 33GB during 25-step generation and 41GB during 4-step LoRA execution.

2. Identifying Active LoRA Layers

Using the LoRA Block Weight node, I selectively enabled subsets of transformer layers and measured variance against full-layer LoRA outputs.
Matching the author’s test cases, I evaluated layers 24–41, 30–41, 36–41, and 0–23.

Active LayersAuthor (5090): Comp / Edge / Text / ColorM1 Max: Comp / Edge / Text / ColorVisual Assessment
Layers 24–412.8 / 0.10 / 0.08 / 0.73.06 / 0.10 / 0.08 / 0.69Virtually identical to all layers
Layers 30–416.2 / 0.27 / 0.18 / 2.25.95 / 0.27 / 0.19 / 2.34Intact; minor detail shifts
Layers 36–4112.4 / 0.70 / 1.82 / 9.412.82 / 0.73 / 1.77 / 9.97Double outlines and blurring
Layers 0–2312.9 / 0.77 / 4.04 / 13.213.29 / 0.78 / 4.07 / 13.72Severe degradation (same as no LoRA)

The author’s reference data was measured on a compressed variant (moss_image_sft_4_steps_lora_v2_600_avg_rank_72_bf16).
Running the official full v2.0 adapter directly on M1 Max yielded metrics nearly identical to the author’s RTX 5090 numbers across all four criteria.

Layers 24–41 only
Fox with LoRA on layers 24-41, almost identical to full LoRA
Layers 36–41 only
Fox with LoRA on layers 36-41, showing double outlines
Layers 0–23 only
Fox with LoRA on layers 0-23, corrupted into unrecognizable blur

The effective updates of the 4-step LoRA are concentrated heavily in layers 24–41. The weights in layers 0–23 can be discarded without noticeable impact on output fidelity.

3. LoRA Rank Reduction with LoRA Resize

The LoRA Resize node computes SVD to compress an adapter down to a target rank and exports a new .safetensors file.
I created compressed versions at rank 8 and rank 2 from the original rank 128 adapter.

FileSizeProcessing TimeRetained Variance (Mean)
Original LoRA (Rank 128)1.19GB——
Rank 833.1MB7.3s68.3%
Rank 28.3MB6.0s56.3%

Retained variance reflects how much original weight difference is captured in the truncated rank, averaged across all layers. At rank 2, only ~56% of total variance is preserved mathematically.

LoRA Resize does not prune specific layer blocks during export.
To reproduce the author’s layer-filtered benchmark, I passed the 8.3MB rank 2 file through the LoRA Block Weight node, applying it exclusively to layers 24–41.

Image variance was measured against full-layer rank 128 generations:

ConfigurationPhoto / Illustration (Comp / Edge / Text / Color)Typography (Comp / Edge / Text / Color)
Rank 8, All Layers2.71 / 0.10 / 0.06 / 0.64—
Rank 2, All Layers4.03 / 0.15 / 0.15 / 1.303.59 / 0.08 / 0.19 / 0.94
Rank 2, Layers 24–41 only4.22 / 0.17 / 0.16 / 1.333.77 / 0.10 / 0.19 / 0.99

Composition variance remained between 2.7 and 4.2 across all configurations—substantially below the seed variance baseline (16.49).
Visual inspection showed consistent subject placement, lighting, and anatomy. Minor shifts occurred in secondary details, such as the fisherman’s hat style shifting at seed 1002. Ghosting at cafe seed 1001 was faithfully preserved even at rank 2.

In typography tests, all 27 images across the three prompts accurately spelled words such as FRESH BREAD DAILY and NORTHERN LIGHTS.

As the author noted in RESULTS.md, the weight changes in layers 24–38 (v, ffn_up_gate, ffn_down) have over 99% of their variance concentrated along rank 2–5 directions. Because the core dynamics are predominantly low-rank to begin with, compressing to rank 2 retains virtually all task-relevant modifications.

Generation latency remained identical at ~47 seconds for both 1.19GB and 8.3MB files. Because ComfyUI pre-merges LoRA weights into diffusion model parameters prior to sampling, LoRA rank does not alter inference speed.

4. Merging into an int8 Standalone Model

Finally, using the MOSS Model Saver node, I baked the rank 2 LoRA on layers 24–41 directly into the base SFT diffusion weights.
I selected the “int8 ConvRot (Comfy-Org layout)” format, matching official Comfy-Org quantized releases. Linear and attention projections in the diffusion model are converted to 8-bit integers (int8), while remaining components stay in bf16. Note that this creates a standalone diffusion model; the text encoder and VAE remain external standard components loaded alongside it.

Serialization took 3.8 seconds, yielding a single 5.41GB .safetensors file.
Moving this file to models/diffusion_models allowed standard ComfyUI Load Diffusion Model workflows to load it directly without any LoRA loader nodes. Running the same 18 prompts with euler, simple, 4 steps, and CFG 1 produced:

SFT + Original LoRA (1.19GB)
Bakery wooden sign generated with full 4-step LoRA
int8 Standalone Model (5.41GB)
Identical sign generated with int8 standalone model
SFT + Original LoRA (1.19GB)
Fisherman portrait generated with full 4-step LoRA
int8 Standalone Model (5.41GB)
Fisherman portrait from int8 standalone model
ComparisonPhoto / Illustration (Comp / Edge / Text / Color)Typography (Comp / Edge / Text / Color)
int8 Standalone vs. Rank 2 (Layers 24–41)2.62 / 0.10 / 0.11 / 0.702.28 / 0.03 / 0.12 / 0.67
int8 Standalone vs. Original LoRA4.41 / 0.19 / 0.14 / 1.483.97 / 0.11 / 0.22 / 1.03

The int8 standalone model outputs matched the original LoRA compositions closely, and typography in all 9 text images was rendered correctly.

MetricSFT + Original LoRAint8 Standalone Model
Storage FootprintSFT 10.2GB + LoRA 1.19GB5.41GB (Single File)
Model Weight Load Memory9693.73MB5162.83MB
Peak Physical RAM During Sampling41GB31GB
Median Time per Image46.93s51.14s

The int8 standalone model reduced peak physical memory by ~10GB.
However, generation time increased by ~9% (46.93s to 51.14s). Checking the ComfyUI console log showed empty Native ops: and emulated ops: int8_tensorwise. On Apple Silicon MPS, int8 matrix operations run via software emulation rather than hardware-accelerated integer tensor cores. The author similarly observed slower int8 inference on RTX 5090 under SageAttention.

Benchmark Summary

ConfigurationModel SetupLatency (1 Image)Peak Physical RAMKey Takeaway
SFT, 25 stepsSFT 10.2GB574.65s33GBBaseline. Unusable for interactive exploration on Mac
SFT + 4-step LoRASFT 10.2GB + LoRA 1.19GB46.93s41GB~12x speedup to 47s, but memory reaches 41GB
SFT + Rank 2 LoRA (Layers 24–41)SFT 10.2GB + LoRA 8.3MB47.38s41GB8MB adapter achieves identical quality and speed
int8 Standalone Model5.41GB single file51.14s31GBSaves 10GB RAM at the cost of ~9% slower generation

HaDeS MOSS Suite enables MOSS-Image to drop from 10 minutes to ~50 seconds per image on an M1 Max.
The author’s findings—that the 4-step LoRA operates predominantly on layers 24–41 and compresses cleanly to rank 2 without image degradation—held true on Apple Silicon.

References