"Combined Workflow" v8.2 (20260531)

This "Combined Workflow" (about 2MB and over 600 nodes) is a ComfyUI txt2img workflow that performs SDXL, Pony, Illustrious, Flux1D, Qwen and ZImageTurbo/Base, Anima and Flux2Klein generations with an optional prompt extension using LLMs/Ollama and Wildcards processing.

Note: the workflow is not fast, on a 4090, a batch of two image can take 7 minutes.

It will generate an upscaled 16MP image as the final result while staying as close as possible to the original generation and produce CivitAI compatible metadata for each stage of the image generation.

- Stage 1: Generate the regular image using [Detail Daemon](https://github.com/Jonseed/ComfyUI-Detail-Daemon) sampler, pass it to a selector (can be bypassed for batch generation)

- Stage 2: Upscale to 4MP using a model

- Stage 3: Use "Ultimate SD Upscale" to redefine the components of the 4MP image using the original model and LoRAs' specific characteristics. An optional Flux.1D resampler is available followed by a Cleanup stage, then Faces, Hands and Eyes Detailer are then used on the resulting image.

- Stage 4: That result is sent to SeedVR2 or HighresFix to generate the final 16MP image and a color matching step is performed to make it as close as possible as the initial upscaled image.

The workflow contains a READ ME FIRST section that details some information about how it came to be, what it does and how to use it. Please refer to it for more details.

FYSA: list (and count) of used custom nodes:

   1  "cg-image-filter",
 221  "comfy-core",
   4  "comfy-image-saver",
   2  "comfy-mtb",
  18  "comfyliterals",
   1  "comfyui_controlnet_aux",
  12  "comfyui_essentials",
   9  "comfyui_llm_party",
   2  "comfyui_ultimatesdupscale",
  12  "comfyui-crystools",
  61  "comfyui-custom-scripts",
   8  "comfyui-detail-daemon",
  51  "comfyui-easy-use",
   5  "comfyui-fbcnn",
  69  "comfyui-image-saver",
  41  "comfyui-impact-pack",
   1  "comfyui-inspire-pack",
  35  "comfyui-kjnodes",
   7  "comfyui-lora-manager",
   6  "comfyui-ollama",
   4  "comfyui-qwenvl",
   1  "comfyui-resolution-master",
  13  "comfyui-rmbg",
  75  "rgthree-comfy",
   3  "seedvr2_videoupscaler"

Previous releases:

• v8.1 (20260504): Updated to work with updated cg-image-filter.

• v8 (20260308): Added QwenVL as a local LLM (01a, 01b, 01z) + Added QwenVL based "image to prompt" extraction (01c) + Added a new post-positive prompt generation optimization (01z) + Added Detail Deamon for all Samplers and a couple resampler + Added ZimageBase and Anima

• v7 (20260216): Added external LLMs capability + Added an "Automatic" system prompt with model specific rules + Added an optional Flux1.D Resampler (S3, 05a03) + Added a HiresFix "Cleanup" node right after the Resampler steps (S3, 05b01) + Split S3 into its core components (the largest block on the workflow) with many ways to compare from step to step + Added many additional "notes" to explain how to make better use of the workflow

• v6 (20260131): Improved Stage 3: Resampler (2-pass: Regeneration + Details enhancement) and Detailers to follow original conditioning + Continued improvement to "Ollama" logic

• v5 (20260124): Added "Advanced" Ollama prompt + Added a new "LoRA randomizer" group + Implemented SEGS for Detaillers using "small" Face/Eyes/Hands selection logic

• v4.1 (20260118): included setup requirements (diffusion models as checkpoints + Ultralytics required setup) in "READ ME FIRST" section + Changed to a common resolution selector

• v4 (20260111): Addition of alternate samplers for Qwen and Z Image Turbo + removal of node failing to install on new Comfy installation + extended documentation: muted nodes-chain need to be manually selected

• v3.1 (20251231): Hotfix for face/hand size

• v3 (20251230): Additional detailers tweaks + alternative models for refiner/detailer steps

• v2 (20251228): Trigger word selection + Detailers tweak + Usage clarifications

• v1 (20251226): Initial release

Work-in-Progress release:

This workflow(Pony, SDXL, Illustrious only for now) is a Work-In-Progress combination, testing and tweaking of various elements from other workflows to generate an upscaled 16MP image as the final result while staying as close as possible to the original generation as generate CivitAI metadata for each stage of the image generation.

- Stage 1: Generate the regular image, pass it to a selector (can be bypassed for batch generation)

- Stage 2: Upscale to 4MP using HiResFix

- Stage 3: Use Ultimate SD Upscaler (No Upscale) to redefine the components of 4MP image using the original model and loras' specific characteristics. Faces and Eyes Detailer are then used on the resulting image.

- Stage 4: That result is sent to SeedVR2 to generate the final 16MP image.

There are many nodes involved in this workflow. Because of that I made use of multiple subgraphs to keep the workflow organized and easy to navigate.

Groups exists as organizational structure for the entire process and follow the Stage numbers.

It "works for me" but it might not be the best way to do it. Feedback is welcome.

PS: Despite my best effort, I still do not know how to have the "Nodes" used show on each image's page on CivitAI -- if someone knows how to please let me know.

Older releases:

The workflows I use with my Wildcards (see my account for those).

Within the zip is a README.md that explains the various use cases:

• The SDXL, Illustrious and Pony compatible workflow is an extension of @DigitalPastel 's "Smooth Workflow" https://civarchive.com/models/1598938/smooth-workflow-txt2img

• The txt2img Flux and Qwen workflow is an extension of @AlexLai 's "Atomix Workflow" https://civarchive.com/models/878828/atomix-flux-txt2img-workflow

• The Z-Image Turbo workflow is an extension of @MrSmith2025 's "Z-Image Turbo - T2I Workflow + Detailer + SeedVR2 Upscaler + Color Grading" https://civarchive.com/models/2174733/z-image-turbo-t2i-workflow-detailer-seedvr2-upscaler-color-grading

• Each workflow can work with Ollama (if available) to create a multi-line narrative. They use Ollama as part of the workflow, which might be a VRAM resource constraint (setting "keep alive" to 0 will unload the model immediately after the request finishes). The conversion model should be instruction-tuned.