AI-KSK

Posted on Nov 1

Five MOSS ComfyUI Workflows: Baselines, Comparisons, Layout, and Color Control

#comfyui #imagegeneration #workflows #experimentation

A comparison graph is useful only if you know what changes between its branches. Giving two samplers the same prompt does not make the resulting experiment a test of one variable. An added LoRA, a different guidance path, or an empty negative prompt can change what the comparison actually means.

This MOSS collection contains five separate ComfyUI workflows: a color-control LoRA experiment, a three-branch sampling comparison, a Base/SFT comparison, a structured-layout prompt example, and a simple SFT text-to-image starting point. Together they contain 70 nodes, including five documentation notes. They are separate JSON files, not five presets silently merged into one graph.

The following walkthrough is based on their saved nodes, values, and connections. It does not establish that the workflows have run successfully on a particular installation. The accompanying transcript describes the creator's image-comparison experience; that does not provide a new controlled benchmark or prove every claim about this exact set of exported graphs.

Shared preparation

All five graphs select qwen3.5_9b_bf16.safetensors through a CLIP loader with type moss_image, and flux2-vae.safetensors through a VAE loader. They configure EmptyFlux2LatentImage for width 1024, height 1024, and batch size 1. These are latent-input settings, not independently verified image outputs.

The common SFT model filename is moss_image_sft_int8_convrot.safetensors. Workflow 03 also selects moss_image_base_int8_convrot.safetensors. Other required weights are specific to the LoRA branches described below. The collection does not bundle weights, a tested dependency lockfile, or verified upstream download URLs. Resolve the matching node implementations and asset versions before queueing; do not treat filenames as proof of training method, quantization behavior, or licensing.

These are text-to-image examples. No LoadImage node is present in the five supplied graphs. Their promotional covers are not required inference inputs, and a cover's subject need not match the prompt currently stored in its corresponding JSON.

Several sampler arrays contain a control string in addition to the listed widget names. Rather than inventing a positional schema, this article distinguishes the authors' step/CFG descriptions in the notes from directly named controls and graph connections. Inspect the sampler UI against the installed node version before relying on a parameter mapping.

Workflow 05 — SFT text-to-image starting point

Purpose and assets. Start here if you want to inspect one straightforward generation path before comparing branches. It uses the common SFT model, encoder, and VAE, with SkipLayerGuidanceDiTSimple between the model loader and sampler.

Current configuration. The note describes the intended preset as 50 steps, CFG 3, Euler, simple, and SLG22. The SLG node itself stores double_layers="", single_layers="22", start_percent=0, and end_percent=1. The latent is configured for 1024×1024, batch 1. The positive text describes a misty mountain valley at sunrise; it is an input prompt, not a description of a verified output.

Use it. Load the graph, resolve dependencies, and edit positive text at node 6. Review negative text at node 7 and canvas controls at node 8. Inspect sampler 9 in the matching UI, then run only after those checks. Decode node 10 feeds save node 11, whose prefix is N01_SFT_50步_入门文生图_11.

Limitations. The note mentions 2K experiments and a separate preset document. Neither establishes a 2K default or an additional included file. No output, runtime, or compatibility result is claimed here.

Workflow 03 — Base versus SFT, with shared inputs

Purpose and assets. This graph is for comparing two selected model files, not proving that one model family always wins. In addition to the shared assets, it needs the Base filename listed above.

Current configuration. Nodes 7 and 12 load Base and SFT filenames respectively. Each model passes through its own SLG node. Both branches receive the same positive conditioning from node 4, negative conditioning from node 5, latent from node 6, and VAE from node 3. Their saved sampler arrays are identical: [777, "fixed", 50, 3, "euler", "simple", 1]. The note labels both branches as 50-step/CFG3/SLG22.

Use it. Edit the shared teapot-scene prompt at node 4 once, then confirm that both branches still receive it. Check the two model selections and corresponding SLG and sampler settings before running. Save nodes 11 and 16 provide separate destinations; label real results by branch rather than guessing from appearance.

Limitations. Matching stored arrays and shared links support a configuration comparison, not equal measured execution behavior. A pair of selected filenames does not demonstrate training history, overall quality rankings, or controlled results across many prompts.

Workflow 02 — Three named sampling presets, not one changed variable

Purpose and assets. This graph organizes the source's 50/8/4 comparison into three branches sharing one SFT model loader. The additional weights are MOSS_DMD_8STEP_QKV_FORGE_FIX.safetensors and moss_image_sft_4step_v2.0.safetensors.

Current configuration. The first model branch passes through SLG node 9. The other branches pass through LoRA nodes 14 and 19, each with strength_model=1. All three share positive conditioning and a configured 1024×1024 latent. However, the first branch receives the populated negative text from node 5, while the two LoRA branches receive empty negative text from node 6.

The note describes the first preset as 50 steps/CFG3 and the other two as 8 or 4 steps/CFG1. Their raw sampler arrays also differ. It would therefore be incorrect to write that only the step count changes: the model-processing paths and negative conditioning differ too.

Use it. Update the shared greenhouse-scene prompt at node 4. Confirm the two distinct LoRA selections and both negative-text sources. Inspect sampler nodes 10, 15, and 20 in the installed UI, then associate actual saved files with nodes 12, 17, and 22.

Limitations. The note explicitly leaves runtime acceptance pending after a structural simplification. The branch labels are not measured speed ratios. If you want a controlled experiment, record every changed setting, hardware details, elapsed time, and actual outputs; this export does not supply those measurements.

Workflow 04 — Structured layout as text input

Purpose and assets. This uses the shared SFT path but changes the form of the positive prompt. It does not add a visual bounding-box editor or a dedicated layout-validation stage.

Current configuration. Node 6 stores a JSON string inside CLIPTextEncode. Its fields include summary, items, bbox_coord_space="abs_centralized", and bbox_format="yxyx". The example requests a split-panel travel poster with the text EXPLORE and a blue mountain illustration. Its text box is [-100,-480,100,-40]; the illustration box is [-300,80,300,460].

Use it. Edit the JSON string at node 6 while retaining valid JSON syntax. Update each item's description, literal text, and box consistently. Review the canvas at node 8, then inspect the generated result after an actual run. Save node 11 uses prefix N04_JSON_Layout_结构化布局_11.

Limitations. The coordinate-space and format names are saved prompt fields, not proof of pixel-accurate layout behavior. A syntactically valid JSON string does not guarantee that text will be rendered correctly or that every requested box will be obeyed. Do not present the prompt's requested poster as a verified result.

Workflow 01 — Color-depth LoRA experiment

Purpose and assets. This adds moss_image_color_depth_control.safetensors to the SFT model path before SLG. It is a focused LoRA experiment, not the low-step acceleration workflow above.

Current configuration. Loader 2 feeds LoRA node 3, which stores strength_model=0.8; its output feeds SLG node 4 and sampler 10. The positive prompt at node 7 describes a rain-soaked neon street with magenta and cyan reflections. Decode node 11 connects to save node 12, with prefix N05_色彩深度_LoRA控制_12.

Use it. Confirm the exact LoRA file and strength, edit the positive prompt at node 7, and review negative text and latent settings. If testing changes in strength, record the chosen values and keep the other intended comparison settings explicit. Evaluate actual results rather than assuming the filename proves a particular visual effect.

Limitations. The note corrects an earlier “distillation LoRA” label. It also says a claimed official provenance was not independently verified. Neither acceleration behavior nor official ownership should be inferred from that old label or note.

Choosing a starting point and recording results

Use workflow 05 for the simplest path, 03 for the paired model-file comparison, 02 to inspect three different presets, 04 for structured prompt input, and 01 for the additional color-control LoRA. These choices describe graph organization, not quality rankings.

For a useful first test, record the exact workflow, installed node and weight versions, prompt, canvas, effective sampler settings, and the output you actually inspected. If a node or asset is missing, resolve that before changing multiple unrelated settings. A save prefix, cover image, or author's favorable observation is not a substitute for a run record.

A reliably matched tutorial URL is not included with this collection: the supplied episode description refers to a different subject. No replacement video or RunningHub URL is guessed from an adjacent folder. Workflow, weight, node, artwork, and generated-content rights must also be checked separately; this collection does not assert a universal license or commercial-use permission.

Episode resources

DEV Community

Five MOSS ComfyUI Workflows: Baselines, Comparisons, Layout, and Color Control

Shared preparation

Workflow 05 — SFT text-to-image starting point

Workflow 03 — Base versus SFT, with shared inputs

Workflow 02 — Three named sampling presets, not one changed variable

Workflow 04 — Structured layout as text input

Workflow 01 — Color-depth LoRA experiment

Choosing a starting point and recording results

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

Top comments (0)