Looking for a flexible workflow (Text only OR Face ID + Body/Style references) that translates a flawed body blueprint into a high-quality generation. Need a JSON link, tutorial, or building help!

[dakusourzo-wq]

Hi everyone,

I really need your help. I have spent the last 9 sleepless nights diving deep into ComfyUI, watching dozens of tutorials on YouTube, and reading through endless guides. Unfortunately, I’m still struggling to build or find a workflow that perfectly meets my goals due to my current hardware limitations.

What I am looking for: I need either a direct link to a .json template, a clear step-by-step tutorial, or active help from this amazing community to build a universal, highly flexible ComfyUI workflow for character generation. I want a setup with a standard Text Prompt and two independent Image Input nodes (using IP-Adapter / Flux Redux / PuLID architecture) that can dynamically switch between text-only and image-guided generation using bypass switches (so the pipeline doesn't crash if an image slot is left empty).

My Hardware Stack & Expectations:

Current GPU: RTX 2060 (6GB VRAM)

Current RAM: 16GB (System pagefile is heavily expanded to 48GB on a fast SSD)

Upcoming Upgrade: Planning to upgrade to an RTX 5060 Ti (16GB VRAM) very soon.

CRITICAL NOTE ON SPEED: Speed is completely secondary right now. I don't care if a single generation takes 15–20 minutes or more on my RTX 2060. The main priority is for the workflow to be rock-solid and functional.

EVEN IF MY CURRENT GPU CAN'T HANDLE IT: If this dual-reference logic is fundamentally too heavy for 6GB VRAM right now, please help me build/find it anyway! I want to have this exact workflow ready so I can use it the second my new GPU arrives.

The Core Requirement & Image Logic (Crucial Nuance):

The workflow needs to handle three modes: Mode 1 (Text Only), Mode 2 (Text + Face Image 1), and Mode 3 (Text + Both Images).

Here is the exact logic I need for Mode 3, which is the main puzzle for me:

Image Input 1 (Face & Style Reference): A close-up, ultra-realistic photo of a person's face. This image defines the facial identity, realistic skin texture, lighting, and overall high-quality style.

Image Input 2 (Body & Proportions Reference): An image of a body/outfit. Crucially, this image might be low-quality, have poor anatomy, or look less realistic. However, it contains the "blueprint" of the body: whether the person is tall or short, fat or thin, has long or short legs, an hourglass figure or a straight waist, and what clothes they are wearing.

The Output: If I type a prompt like "A beautiful girl sitting in a Parisian restaurant" and provide both images, the model must use Image 2 only as a structural guide for the body proportions and clothing silhouette. It must completely ignore the low-quality or flawed style of Image 2. Instead, it must organically redraw that exact body type and outfit from scratch, upscaling it to the ultra-realistic style, lighting, and correct anatomy derived from Image 1.

Essentially, it needs to translate a flawed body reference into a biologically correct, highly realistic full-body generation, matching the identity of Image 1, all natively from noise (no post-processing face-swaps like ReActor).

My Question:

Can someone share a link to an existing workflow on Civitai/OpenArt, a tutorial, or guide me on how to chain the adapters/nodes properly to achieve this anatomical translation and 3-in-1 flexibility?

Should this be built on Flux (using highly optimized GGUF Q4 models / Flux-Schnell at 4 steps to survive) or should I map this out on an SDXL multi-IP-Adapter setup first?

After 9 days of trial and error, any shared JSONs, custom node suggestions, or building advice would mean the world to me. Thank you so much!