Hugging Face's logo

AI-KSK
/
moss-image-five-workflows-324

Model card Files
xet
 Community

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Five MOSS Workflows for ComfyUI

One collection, five separate workflow JSON files: color-control LoRA, three-preset comparison, Base/SFT comparison, JSON layout input, and a simple SFT starting point. These files contain 70 nodes in total, including five notes. This is a workflow collection, not a model-weight release or a completed benchmark.

Contents

Code / purpose Workflow Promotional cover
01 — MOSS Color-Depth LoRA Experiment JSON Cover
02 — MOSS Three-Preset Comparison: 50 / 8 / 4 JSON Cover
03 — MOSS Base and SFT: Shared-Input Comparison JSON Cover
04 — MOSS JSON Layout: Structured Text Prompt JSON Cover
05 — MOSS SFT: A Simple Text-to-Image Starting Point JSON Cover

The JSON copies retain the original bytes. Covers are promotional assets, not verified inference results or default inputs. All five graphs use empty latent inputs rather than LoadImage nodes; no reference picture is required by these exported graphs.

Shared prerequisites

The common loaders select moss_image_sft_int8_convrot.safetensors, qwen3.5_9b_bf16.safetensors (CLIP type moss_image), and flux2-vae.safetensors. Workflow 03 additionally selects a Base file; workflows 01 and 02 add the LoRAs listed below. All five configure a 1024×1024 latent with batch size 1. Actual output dimensions and success have not been tested here.

Resolve compatible ComfyUI nodes, including the MOSS loader/encoder configuration, EmptyFlux2LatentImage, and SkipLayerGuidanceDiTSimple. A tested version lockfile, verified upstream download URLs, hardware measurements, and weight licenses are not supplied. Do not install an arbitrary similarly named asset or infer implementation details from filenames.

Sampler arrays contain extra control values beyond the listed widget names. Step/CFG descriptions below are attributed to the source notes where necessary; confirm their interpretation in the matching UI.

Per-workflow setup

01 — MOSS Color-Depth LoRA Experiment

Inspect a model-only color-control LoRA inserted before SLG, rather than treating it as a low-step distillation preset.

Saved configuration: Model loader 2 feeds LoRA 3 at strength_model=0.8, then SLG 4 and sampler 10. The configured latent at node 9 is 1024×1024, batch 1.

Run preparation: Select the matching SFT model and moss_image_color_depth_control.safetensors. Edit the neon-street prompt at node 7 and review negative text at node 8. Confirm the LoRA strength and the sampler settings in the matching UI before running. Decode 11 feeds save node 12; record actual results when comparing strengths.

Additional weights: moss_image_color_depth_control.safetensors

Limits: The note corrects an earlier distillation-LoRA label and explicitly leaves claimed official provenance unverified. Neither acceleration nor a particular color improvement is established.

02 — MOSS Three-Preset Comparison: 50 / 8 / 4

Inspect three named sampling presets sharing an SFT loader, while keeping their different model-processing and conditioning paths explicit.

Saved configuration: Loader 8 branches through SLG 9, DMD LoRA 14, and Turbo LoRA 19. Both LoRA strengths are 1. The SLG branch uses negative text 5; both LoRA branches use empty negative text 6. Positive text 4 and latent 7 are shared.

Run preparation: Provide MOSS_DMD_8STEP_QKV_FORGE_FIX.safetensors and moss_image_sft_4step_v2.0.safetensors in compatible versions. Edit the greenhouse prompt once at node 4. Check both negative-text inputs and all three model paths. Inspect samplers 10, 15, and 20 before queueing; distinguish actual saved outputs from nodes 12, 17, and 22.

Additional weights: MOSS_DMD_8STEP_QKV_FORGE_FIX.safetensors, moss_image_sft_4step_v2.0.safetensors

Limits: The note describes 50-step/CFG3 and 8- or 4-step/CFG1 presets, but steps are not the only changed variable: LoRAs, SLG, and negative conditioning differ. No measured speed ratio, quality ranking, or runtime acceptance is supplied.

03 — MOSS Base and SFT: Shared-Input Comparison

Compare two selected model files through separate branches that share text conditioning, a latent input, and a VAE.

Saved configuration: Model loaders 7 and 12 select Base and SFT filenames. They feed separate SLG nodes 8 and 13. Samplers 9 and 14 have identical saved arrays [777, "fixed", 50, 3, "euler", "simple", 1]. The note labels both branches 50-step/CFG3/SLG22.

Run preparation: Provide moss_image_base_int8_convrot.safetensors as well as the common SFT file. Edit the teapot-scene prompt at node 4 once; confirm shared negative text 5, latent 6, and VAE 3. Inspect each model and sampler in the matching UI, then label actual results from save nodes 11 and 16 by branch.

Additional weights: moss_image_base_int8_convrot.safetensors

Limits: Identical saved arrays are configuration evidence, not proof of identical runtime behavior or broad quality superiority. Filenames do not establish training history or a universal Base-versus-SFT conclusion.

04 — MOSS JSON Layout: Structured Text Prompt

Explore a structured layout request placed directly in a text-encoding node, not a visual box editor or a verified layout engine.

Saved configuration: Node 6 contains a JSON string with summary and items, bbox_coord_space=abs_centralized, and bbox_format=yxyx. It requests EXPLORE text and a blue mountain illustration. Latent node 8 stores 1024×1024, batch 1.

Run preparation: Edit the JSON string in positive-text node 6. Keep JSON syntax valid while updating each item description, literal text, and box consistently. The stored example boxes are [-100,-480,100,-40] and [-300,80,300,460]. Check the latent and sampler UI, then inspect actual text and placement after running; save node 11 is the output destination.

Additional weights: None beyond the common selections.

Limits: These coordinate fields are saved prompt content, not implementation documentation or proof of pixel-accurate placement. There is no dedicated bounding-box editor or JSON-validation node in this graph. Requested poster content is not an observed result.

05 — MOSS SFT: A Simple Text-to-Image Starting Point

Start with one SFT-to-SLG sampling path before examining the collection’s more complex comparison branches.

Saved configuration: Loader 2 feeds SLG 3 and sampler 9. SLG stores double_layers="", single_layers="22", start_percent=0, end_percent=1. Latent 8 stores 1024×1024, batch 1. The note describes a 50-step/CFG3/Euler/simple preset.

Run preparation: Edit the misty-valley prompt at node 6, review negative text 7, and inspect the latent controls at node 8. Confirm the effective sampler configuration in the installed UI rather than treating an unnamed control value as a schema. Decode node 10 feeds save node 11.

Additional weights: None beyond the common selections.

Limits: The note’s mention of 2K tests is not a 2K default. A separately mentioned preset document is not supplied in this package. No output, GPU requirement, or compatibility result is established.

Running and troubleshooting

Import one JSON at a time, resolve missing nodes and compatible weights, edit the appropriate prompt, and review connected inputs before queueing. Record effective settings and distinguish each actual output by its save prefix. Wrong filenames or unresolved nodes should be corrected before changing unrelated sampler settings. For a comparison, record every changed branch—not just the intended variable.

Evidence, tutorials, and rights

Only static JSON inspection is included. There are no new inference outputs, GPU measurements, speed ratios, or verified quality rankings. Save nodes and cover art are not execution evidence. Tutorial links are omitted because the provided episode description does not reliably match these materials; no adjacent episode URL is substituted.

The supplied workflow notes retain source labels M14, M06, M03, M08, and M02 for the five derivatives. They are provenance notes, not license grants. Verify rights for the workflow, each model and LoRA, node code, cover artwork, and generated content separately. No repository-wide license or universal commercial-use permission is asserted.

Episode resources

01 — MOSS-Image｜色彩深度LoRA·光影与色彩控制实验版

02 — MOSS-Image｜50步原版 vs 8步DMD vs 4步Turbo·三档加速横评版

03 — MOSS-Image｜Base vs SFT 双模型同参对决·画质横评版

04 — MOSS-Image｜JSON Layout 结构化布局·图文排版控制版

05 — MOSS-Image｜SFT 50步高质量文生图·新手入门版

Hugging Face: https://huggingface.co/AI-KSK/moss-image-five-workflows-324 · ModelScope: https://modelscope.cn/models/Sunjian520/moss-image-five-workflows-324

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support