iGEN ONE M Nova — Workflow Guide

A modular image generation, editing and finishing workflow for ComfyUI.

This guide covers IGEN_ONE_M_Nova.json, the smaller M edition. Its 25 groups bring together prompting, initial generation, image input, upscaling, reference-based editing, two detailers and final output. All group descriptions and instructions below refer to this M workflow.

The workflow uses three separate Nova repositories:

• ComfyUI Eclipse Nova — Prompt tools, routing, image loaders, the advanced sampler, detailers, image processing and saving.

• ComfyUI Smart Model Loader Nova — Diffusion model loading, CLIP/VAE selection, LoRA stacks, conditioning and the KSampler Pipe used in Upscale.

• ComfyUI SmartLLM Nova — Image descriptions, prompt refinement, tile descriptions and detection.

Nodes marked [SML2] belong to Smart Model Loader Nova. That saved node identifier is intentional. Install the Nova repositories for this workflow.

Start Small

For a first text-to-image run:

1. In Folder, choose the initial image size and set the batch size to 1. The saved file currently uses a batch of 3.

2. In Sampler, choose a compatible model, text encoder and VAE in Smart Model Loader. Check the filenames against your installed files.

3. In Prompt, enable the sources you want. For a simple manual prompt, use Wildcard Processor and turn off Subject and Settings if you do not want their text added.

4. Leave the image-input groups muted. Keep Initial Image Generation enabled and Encode Input Image disabled in the Sampler group's toggle panel.

5. Choose where sampling settings should come from using Use pipe settings in Eclipse KSampler Advanced. See the priority rules below.

6. For the simplest run, mute Upscale, both detailer groups and Image Adjustments, along with any optional editing or watermark groups. The saved file has several processing stages active.

7. Keep Save Image: Final active and queue the workflow.

8. When Image Selector pauses, select the image or images to keep and press Confirm to continue.

The saved workflow contains personal model selections, paths, prompts and processing choices. Treat these as an editable starting configuration.

How the Groups Connect

Most intergroup connections use Set/Get nodes. A Set node publishes a named value; Get nodes retrieve it elsewhere. Get All Active and Any Multi-Switch choose the first available source from an ordered list.

The image path is:

Image input or initial generation → Upscale → optional intermediate Save Image → Flux2 Refine → Qwen Image Edit 2.1 + Face Swap → Krea2 Edit → Tile Upscale → Detailer_1 → Detailer_2 → Image Adjustments → Watermark (Text) → Watermark (Logo) → Save Image: Final.

Mute optional groups to skip them. Later groups then use the previous available result. Group positions help with navigation; the named connections determine the processing order.

Use a group's Fast Mode Toggle for individual features and the canvas group controls to enable or mute a whole optional group. Activating every node in a group can also activate optional branches, so review its toggle panel afterwards.

Nova's collapsible sections keep controls compact. Collapsing a section hides its controls; use its feature toggle to disable the operation itself.

Folder and Initial Resolution

Folder contains Smart Folder, which supplies the output folder, initial width and height, and batch size. The saved size preset is 2:3 at approximately 1 megapixel, aligned to a divisor of 32.

These dimensions are the starting canvas. Later stages can resize independently: Latent upscale between sampling stages, the separate Upscale group, Tile Upscale and Image Adjustments.

The save groups record the initial Folder dimensions in generation metadata. The saved image still has its actual final pixel dimensions. This helps reconstruct the original generation before its upscale passes.

For reproduction, set Folder to the original generation size. Loading sampling metadata does not automatically change the Folder controls.

Image Inputs and Preparation

Image Load

Browse an input/output image or upload your own. Load Image passes the image and selected metadata through its PIPE.

Enable Metadata to inspect embedded generation data. Include Sampler, Positive prompt, Negative prompt and Seed independently:

• Selected sampling values and seed reach the main sampler through Loaded image sampling data.

• Selected prompt text enters prompt assembly when Image Load Metadata is enabled in the Prompt group's toggle panel.

This lets you reuse sampling settings without adding the original prompt, or reuse the prompt with fresh sampling settings.

Load Images / Video Frames

Load Files From Folder supplies images or decoded video frames from your paths. Use Single, Range or Step to choose how much to load. Selection and advancement have explicit controls; use After run → Shuffle where available instead of the old negative-index shortcuts.

The connected Image Selector presents the loaded images for review. Select one or more and press Confirm to continue. Pipe Slicer & Dice keeps the selected images and filenames aligned.

The sampling/prompt metadata route in this M file is wired from the separate Image Load group. The folder/frame selector supplies selected images; it is not wired as another sampling-metadata source.

Remove Background

The BiRefNet branch removes the source background. Its preprocessing component includes model and background/fill controls. Check the preview before continuing.

Crop Image (Custom)

Inset & Crop trims or reframes the selected source. Review the composition before resizing.

Image Resize

Resizes the prepared input using Folder's width and height. Use it before image-to-image sampling when the source should match your initial canvas.

Image to Prompt

Smart LM Loader describes the selected image. The following String Multiline node can filter unwanted terms before the description enters the main prompt.

Choose an image-capable model and a suitable task. The saved branch uses a ComfyUI text-encoder model; SmartLLM also offers registry-backed alternatives. An Ollama setup is needed only when you select an Ollama model.

The sampler's source-image priority is Image Resize → Image to Prompt's passed-through image → Crop Image (Custom) → Remove Background → Load Images / Video Frames → Image Load. Enable the source and preparation steps you intend to use.

Building the Prompt

Prompt

The central group combines its enabled sources:

• Wildcard Processor — Manual text, wildcard files and random choices.

• Smart Prompt: Subject — Structured subject descriptions from editable prompt files.

• Smart Prompt: Settings — Scene, camera, lighting and style descriptions.

• Image to Prompt — The description generated from your source image.

• Image Load Metadata — Prompt text extracted from the loaded image.

• Danbooru Prompt Forge and Prompt Lists — Generated tags or saved prompts.

• Prefix / Suffix — Optional text before or after the assembled prompt.

Use the toggle panel to choose which sources participate. Enabling several combines their text. Negative Prompt provides manual negative text, followed by deduplication.

The Negative Prompt toggle in Sampler decides whether that text is encoded for initial sampling. With it disabled, the workflow uses zeroed negative conditioning.

Danbooru Prompt Forge

Builds prompts from tag categories, required tags, exclusions and custom text. It receives the shared prompt seed. Enable its source in Prompt to include the result.

Prompt Lists

Read Prompt Files reads a text-file prompt library. Choose the starting index and advancement mode, including shuffle. String Multiline filters the result and can pause for review.

A connected, unchanged prompt seed freezes the selection. Change it to advance the seeded selection. The displayed last index identifies the selection used; the starting-index control does not need to change after every run.

Replace the saved file path with your own library before enabling this group.

Prompt Styler

Applies a style to positive and/or negative text. Show Any Stop displays the result for review. Select a style suited to your model and choose which prompt fields it affects.

Prompt Refine

Smart LM Loader rewrites the assembled prompt using the selected task, currently Refine & Expand Prompt. Review the result when exact wording or pose matters.

The final positive prompt prefers Prompt Refine, then Prompt Styler, then the basic assembled prompt. The negative path uses the styler result when available, then the basic negative prompt.

Sampler: Models, Settings and Seeds

Smart Model Loader and LoRAs

The main Smart Model Loader sits inside Sampler. Choose the model, encoder type/files and VAE together. The saved file uses Krea2; changing families may require a different encoder, VAE and conditioning setup.

Enable the desired entries in LoRA Stack and set their strengths. The patched model and CLIP are also published for downstream groups that reuse the main model.

Where Source / integrity contains an AIR or repository reference, Open in Download Manager can retrieve the file. A completed download opened from that loader fills its associated file and integrity fields. Saving or updating a preset is a separate user action.

Which Sampling Settings Are Used?

Use pipe settings on Eclipse KSampler Advanced is the master choice:

• Enabled: supplied PIPE values override corresponding sampler controls. Missing values stay local.

• Disabled: settings and seed come from the sampler panel. The connected model, VAE, conditioning and latent still provide the generation inputs.

For the incoming standard sampling values, this workflow uses:

1. Enabled, available Sampler metadata from Image Load for steps, CFG, sampler and scheduler.

2. Otherwise, enabled sampling settings from Smart Model Loader.

3. Local sampler controls for fields absent from the PIPE.

The incoming seed prefers enabled image metadata, then the dual node's Image seed. These incoming seeds apply when Use pipe settings is enabled.

If changing the displayed steps or CFG has no effect, check the upstream settings. To edit sampling values locally while retaining the connected image seed, disable the loader's Sampler include option and the loaded image's Sampler metadata selection, while keeping Use pipe settings enabled.

Image / Prompt Seed

The dual seed node separates two kinds of variation:

• Image controls diffusion noise for initial sampling.

• Prompt controls the connected Smart Prompt builders, Prompt Forge and prompt-file selection.

Both share a 64-bit toggle and have independent seed controls. Keep Prompt fixed while exploring image variations, or keep Image fixed while changing the connected prompt choices.

Wildcard Processor has its own local seed in the saved file; it is not connected to the shared Prompt output. If Use pipe settings is off, the advanced sampler also uses its local seed.

Standard or Two-Stage Sampling

Eclipse KSampler Advanced receives a PIPE from Initial sampling inputs and returns the decoded image, latent and resolved settings through Initial sampling outputs.

• Two-stage off: normal sampling with steps, CFG, sampler, scheduler and denoise. This is the saved M file's current mode.

• Two-stage on: Stage 1 and Stage 2 have separate step ranges and noise controls. The usual sampling fields describe Stage 1; Stage 2 has its own settings.

• Latent upscale: optionally enlarges the latent between stages. It has no effect while Two-stage is off. The saved choice is bicubic at 2×.

• Stage 2 seed: an explicit value takes priority; otherwise it uses the base seed plus the configured offset.

The workflow leaves model_2 unconnected, so both stages reuse the main model. IO KSampler can accept a compatible second model for a high/low-noise setup, but the intended handoff requires explicit step-range and noise settings.

Before enabling Two-stage, review its ranges together with the effective Stage 1 settings. A longer-schedule recipe is not a universal preset for an 8-step turbo model.

Optional Qwen schedule, DiT skip guidance and Detail Daemon sections always use local controls, even with Use pipe settings enabled. Qwen schedule replaces the normal sigma schedule while active; use it with a compatible model. Detail Daemon is built into Eclipse Nova and needs no separate node package.

Leave Alpha off to output RGB for downstream processing. Tiled decode can reduce decoding memory requirements.

Image Selector

After generation, select which batch results continue and press Confirm. The slicing step keeps the corresponding prompt and conditioning entries aligned.

Confirmation reuses the reviewed results and preserves upstream seeds for that continuation. Use Auto confirm when you want all images to proceed without manual selection.

Working from an Existing Image

Image-to-Image Generation

1. Enable Image Load or Load Images / Video Frames and choose the source.

2. Enable preparation steps as needed, then Image Resize to match the intended initial size.

3. Keep Initial Image Generation enabled in Sampler.

4. Enable Encode Input Image. The encoded source takes priority over the empty starting latent.

5. Start with Two-stage off and choose denoise in the settings source actually controlling the sampler. Lower denoise preserves more of the input.

6. Queue, review and continue.

Loading an image alone does not enable image-to-image sampling. Encode Input Image chooses that path.

Reuse Image Prompts and Sampling

1. In Image Load, enable Metadata and the fields you want.

2. Enable Image Load Metadata in Prompt to include its prompt text. Disable other sources if you want to avoid adding text to the original prompt.

3. Keep Use pipe settings enabled to accept the selected sampling values and seed.

4. Set Folder to the original generation dimensions.

5. For reconstruction from noise, leave Encode Input Image off. Enable it for an image-to-image variation of the loaded pixels.

Skip Initial Generation and Only Post-Process

1. Enable an image source and select the images to process.

2. Turn off Initial Image Generation in Sampler's toggle panel. Leave Encode Input Image off when no initial sampling is needed.

3. Keep the Sampler group's routing helpers active: they forward the loaded image to downstream processing.

4. Enable only the processing groups you want and keep Save Image: Final active.

The generation toggle also disables the initial result selector and sampled-settings branch. The loaded image is forwarded automatically. Upscale remains an independent diffusion pass; mute it too if you want processing without that sampling step.

Upscaling, Editing and Detailing

Upscale

Image Filter Adjustments Advanced resizes/sharpens the initial image, then Smart Model Loader Nova's KSampler Pipe refines it. This group reuses the main model with its LoRAs and the selected initial conditioning.

The saved setup resizes by 1.5× and runs a short, low-denoise pass. Its Use pipe settings is off, so local sampling controls apply. Image Comparer lets you inspect the result and choose which image continues.

This group is separate from the advanced sampler's latent upscale. Enabling both can increase resolution more than once.

Flux2 Refine

A separate Flux2 model refines the current image at low denoise. It has its own loader and ComfyUI custom-sampling chain. Compare the result to judge changes in texture and composition.

Qwen Image Edit 2.1 + Face Swap

Uses its own Qwen Image 2.1 loader and edit conditioning, with editable positive/negative instructions. The current image is the edit source.

For reference-assisted face editing, load Reference Image 2 and enable that reference in the preprocessing controls. Smart Detection and Image Crop by Mask prepare a cropped reference; inspect its preview before proceeding. Describe the intended change in the edit prompt and compare the result.

Select the encoder and VAE required by the Qwen Image 2.1 model. This branch has its own model setup.

Krea2 Edit

An instruction-based editing branch with its own loader, LoRA stack and grounded encoding. The preprocessing component exposes the edit prompt, grounding size, reference strengths and fit mode. Enable a second reference image when needed.

This branch uses comfyui-krea2edit. Configure the matching model/LoRA and edit instruction before enabling it.

Tile Upscale

Prepares the image at the target scale, splits it into overlapping tiles, generates tile descriptions with Smart LM Loader, samples them and decodes/assembles the result.

It has its own model loader and sampling chain. Review tile prompts and previews, then compare the assembled image. Tile size, overlap and target scale affect memory use and seams.

Detailer_1 and Detailer_2

The M workflow has two detailers, currently targeting face and mouth. Change the detection target to work on another region.

Each group detects the region, converts detections to bounding boxes, prepares a mask with SAM2Ultra and mask processing, then converts it to SEGS. SEGS Preview Simple shows the selected regions; Detailer (SEGS/pipe) samples and blends them back. Use Image Comparer to review the change.

The Model Loader toggle controls the model source:

• On: use that detailer's own model, CLIP and VAE. With Use pipe settings enabled, its loader also supplies available sampler settings.

• Off: reuse the main model and CLIP with the main LoRA stack, plus the main VAE. The detailer's own settings PIPE becomes unavailable, so its local sampling controls apply.

You can therefore reuse the main model while choosing separate steps, CFG and denoise for each detailer. Turning off Use pipe settings also permits local sampling values with a dedicated detailer model.

The saved file uses Florence-2 phrase grounding. Inspect the detection and mask when changing targets or models; a label alone does not guarantee every detector can find that region.

Image Adjustments

The combined Image Filter Adjustments Advanced panel handles resizing, sharpening, color/tone and finishing operations. Enable the sections you need and inspect the comparer.

The saved M configuration includes another 1.5× resize here. It can change output dimensions even if you only opened the group to adjust color. Disable Scale to preserve the incoming resolution.

Watermark (Text)

Text Image with FX adds text with font, placement, color, opacity and optional effects. The saved text is © Eclipse Nova; replace it with your own credit before publishing.

Watermark (Logo)

Loads a logo and places it with Image with FX. Adjust scale, placement, opacity and effects; use the logo-processing controls to change its appearance before compositing.

Saving and Generation Metadata

There are two save groups:

• Save Image — An optional checkpoint after initial generation/Upscale, before later edits and detailers. It is muted in the saved file.

• Save Image: Final — Saves the last available result in the processing chain. Keep it active for final output.

The final save searches backward from the logo watermark through text watermark, adjustments, detailers, tile upscale, edit/refine stages and Upscale, falling back to the initial image. Muted stages are skipped.

Both save groups use resolved generation data:

• When initial sampling ran, the main sampler's output supplies the settings it actually used, after loader or image-metadata overrides.

• When initial generation is skipped, the fallback PIPE carries available loader settings and selected image metadata/seed. This is a metadata fallback, not a record of a new sampling pass.

• The main generation summary is retained; later refinement does not replace it with only the last detailer's settings.

• With two-stage initial sampling, ordinary sampler fields describe Stage 2. Steps is that stage's configured step count, not the sum of both executed ranges. Keep the embedded workflow for the complete setup.

• Generation dimensions come from Folder, preserving the initial size when the final image is larger.

Use PNG, Embed workflow and Generation data to retain the workflow and generation parameters in the image. Output folder and filename controls support placeholders such as %basemodel, %seed, %sampler_name and %steps.

Review Stops and Common Questions

The workflow pauses before saving

Check the last node that produced a result. Image Selector waits for your selection and Confirm. Text/image review nodes and Stop (Review) controls can also interrupt intentionally.

Disable Stop where you want an uninterrupted run. On supported nodes, the play button executes the upstream path up to that node for inspection before a full queue. Stop-enabled review nodes let unchanged inputs continue without repeating the same interruption.

Sampling changes are ignored

Check Use pipe settings, the main loader's Sampler include option and Image Load → Metadata → Sampler. The separate Upscale sampler and each detailer have their own setting-source choices.

I only want to change the image seed

Change Image on the dual seed node and keep Prompt fixed. Disable the image's Seed metadata field if it is overriding Image. When the advanced sampler ignores PIPE settings, change its local seed instead.

The result is larger than expected

Check all enabled size-changing stages: Folder, two-stage Latent upscale, Upscale, Tile Upscale and Image Adjustments. Several increases can accumulate.

A model or file is missing

Select your own matching model files and replace saved media/prompt-library paths. The workflow does not include model weights or personal media. A saved name, AIR or repository reference does not guarantee that its source remains available.

A Nova node is reported missing

Install/update the three Nova repositories, restart ComfyUI and reopen the workflow. Check the full missing node name to identify its pack. Some saved IDs, including [SML2], retain established identifiers despite the repository's Nova name.

Additional Node Packages Used by This M File

Alongside the three Nova repositories and core ComfyUI, the saved workflow contains:

• ComfyUI Impact Pack — ToBasicPipe adapters in the detailers.

• ComfyUI LayerStyle — Mask Grow in detailer masks.

• ComfyUI LayerStyle Advance — SAM2Ultra mask preparation.

• ComfyUI BiRefNet — The optional Remove Background branch.

• comfyui-krea2edit — The optional Krea2 Edit branch.

The Qwen branch also uses core Qwen Image 2.1 Cache, so your ComfyUI build must include it. Install the packages referenced by the saved workflow even if their processing groups initially stay muted.