Simpler version of my now retired "Combined Workflow" that removes Stage 3 and 4 steps (resamplers, detailers, 4->16MP upscalers) in favor of a multi-model workflow that generates 4MP results (some of the extra steps are part of the "extra" included in the archive).
A simpler initial multi-model generation and an initial upscaler. A lot of the details on usage are linked into the workflow's "READ ME FIRST".
How to use it:
0. Use the various toggle to quick select (the arrows at the end of the selection will bring you to the node/subgraph directly).
1. Select a Model to generate an image with, adapts its parameters to match your setup.
2. Decide if you want to perform txt2img (enter a prompt and get an image) or img2img (use a prompt and a base image to generate a new image based on both). When using img2img make sure to provide an image, set the expected base image resolution (in MP) and set a denoise.
3. Decide if you want to write your prompt or use use an image and have QwenVL write a prompt (you can still add to the prompt in the main prompting area). When using QwenVL, load an image and select the preset for the type of response you need.
4. Decide if you want to use LLMs to optimize the final prompt before the image generation occurs (01z). When doing so, decide on the type of prompt you want generated: tags, narrative or both.
5. Click "Run", the image generation will start with your selected parameters. Depending on the batch_size you will have a set of image generated. The final prompt is viewable. After the initial generation, a simple HiResFix will upscale the image to 4MP.
Also in the archive are the extras (with their matching version number if changed).
The extras add some other features from the original workflow:
detailer: Stage3's detailers made generic detailer with manual mask (and drawing for inpainting) selection.
detailer_sam3: Stage 3's detailers with automatic detection (SAM3 based) to automatically detail the mask found with the provided prompt.
resampler: Stage3's Initial and Detailer resamplers.
Flux_resampler: Stage3's Flus resampler.
Description
- Added optional model randomization: replaced checkpoint/UNet loaders with LoRA Manager loaders (requires version 1.2.2 of LoraManager)
- Added optional LoRA randomization
- Expanded prompt enhancement: choose between local QwenVL, external/API models through ComfyUI LLM Party, or Ollama.
- Reworked the positive-prompt pipeline: separate LoRA Manager Pre-Prompt, Main Prompt, and Post-Prompt fields. Only the main prompt and optional image-derived description go through the LLM; trigger words, pre/post qualifiers, embeddings, and LoRA names are assembled afterward.
- Consolidated wildcard prompting into LoRA Manager.
- Rewritten LLM instructions for Narrative, Tags, and Narrative + Tags modes, emphasizing preservation of subject counts, actions, composition, and explicit details while reducing invented content.
- Added explicit "THEME=..." handling and refined weighted-tag handling.
- Revised image-to-prompt instructions to focus on visible details and avoid unsupported assumptions.
- Simplified img2img loading/resizing.
- Cleaned up generation subgraphs and wiring: removed unused outputs, switches, and inherited parameters.
- Expanded workflow notes, including a model-settings reference table, randomization guidance, prompt assembly explanation, and LLM setup instructions.