Have you ever noticed that anime and illustrated checkpoints have a massively better variety of poses, dynamic angles, and framing than realism models? If you try to prompt those same complex poses directly in a realistic checkpoint, the model often fights you or generates flat, boring compositions.
This ComfyUI workflow solves that by using an existing illustration as a structural foundation, cleanly translating those dynamic poses into high-fidelity, realistic photography.
It features automated prompt enrichment, anti-style bleeding filters, and a streamlined, high-contrast two-stage upscaling refinement pass.
🛠️ Key Features & How It Works
1. Smart Automated Prompting (Florence-2 + Regex Filtering)
Instead of forcing you to manually tag everything your input image contains, the workflow passes your image through Florence-2-large-PromptGen to analyze the scene automatically.
The Problem: Florence-2 will naturally describe the image as an "anime style illustration" or "drawing," which bleeds back into the sampler and ruins the realistic effect.
The Solution: The workflow passes the description through a RegexReplace node, automatically stripping out words like anime, illustration, drawing, digital, and anime-style before sending it to the prompt encoder.
Prompt Concatenation: It automatically glues your custom character info to Florence’s cleaned description, along with built-in quality modifiers. In this updated version, the prompt base has been hardcoded with explicit directives for , proportionate head, bright colours to combat typical stylized anatomical skew and washed-out generations.
VERY IMPORTANT:
For best results add in the prompt for the character you want to generate into the "concatenate Text" node, after the quality modifiers (masterpiece, best quality, score_9, score_8, score_7, score_6, ((realistic photo)), proportionate head, bright colours) this will ensure all the details are captured correctly. So if you want to generate best girl, like in my title photo, you should use:
"masterpiece, best quality, score_9, score_8, score_7, score_6, ((realistic photo)), proportionate head, bright colours, souryuu asuka langley\(neon genesis evangelion\), ginger hair, blue eyes, red hair ornament, yellow dress"
2. Streamlined Pure Dual-Stage Generation
Stage 1 (Base Pass): Takes your scaled input image, encodes it, and runs a base pass using the Sam ANIMA Realistic Turbo model coupled with its dedicated Turbo LoRA. This handles the heavy lifting of morphing the art style while keeping the exact pose and framing.
Stage 2 (High-Res Refinement): Upscales the Stage 1 output using RealESRGAN_x2plus. The upscaled pixels are fed straight back into a VAE encoder for a final, ultra-clean KSampler refinement pass.
Important: High-Contrast Saturation Fix (Latent Multiply)
A common issue with multi-stage img2img workflows is that colors can begin to look washed out, muted, or gray during the second pass. To aggressively counter this, the workflow routes the upscaled latents through a LatentMultiply node set to 1.42 by default. This delivers a potent bump to latent contrast and color saturation right before the final sampling pass, keeping your results incredibly punchy and vivid.
⚙️ Core Parameters & How to Tweak Them
Stage 1 Denoise Strength: Default: 0.42. If your source character has massive, highly stylized anime eyes or extremely exaggerated proportions, adjust this primitive node higher (up to 0.63) to let the model morph the features into realistic anatomy. For closer-to-life proportions, it can drop down near 0.36. Don't sweat it if the first pass looks slightly off, Stage 2 will continue to pull it together.
Stage 2 Denoise Strength: Default: 0.40. Handily exposed to a dedicated PrimitiveNode in this version, you can now directly control how much detail the second pass rewrites. Keep it low to preserve structural data, or bump it up if you want the high-res pass to add more micro-textures and skin realism.
Latent Multiplier: Default: 1.42. If you find your final images look overly fried or hyper-saturated, scale this back down toward 1.2 or 1.0. If they need more pop, leave it right at 1.42.
Batch Processing: Fully optimized for batching. The LoadImage section uses a primitive node set to Increment/Wrap, meaning you can queue up an entire folder of source poses, hit run, and let it churn through them automatically.
💡 Pro-Tips for Best Results
🌟 The Double-Pass Trick: If you have an input image with incredibly exaggerated anime proportions (massive eyes, tiny jaw) and a single run still leaves them looking a bit uncanny, simply save your output and run it through the workflow one more time as the new input image. The second lifecycle completely normalizes the facial features into realistic anatomy.
📝 Character Targeting: While Florence-2 does a fantastic job recognizing clothes and backgrounds, you will get the absolute best results if you use the prompt text field to explicitly define who your characters are (hair color, ethnicity, specific traits) so the model knows exactly how to interpret the shapes.
🧩 Model Requirements
To run this out of the box, make sure you have these in your ComfyUI directories:
UNET Checkpoint: samANIMARealistic_turboV21.safetensors
CLIP Text Encoder: qwen_3_06b_base.safetensors
VAE: qwen_image_vae.safetensors
Florence-2: MiaoshouAI/Florence-2-large-PromptGen-v2.0
Upscale Model: RealESRGAN_x2plus.pth