Workflow is able to +4K Text-to-Image and also Multi-Reference Image-to-Image.

UPDATES: v3.0 - Removed SDE node and replaced with 3 new nodes. Removing SDE and adding the others gives a much more detailed natural look to skin, while a bit softer the tradeoff was worth it IMO, also seems to have better prompt adherence.

v2.0 on left, v3.0 on right

Clown-Shark Workflow v3.0:

It'’s got two sampler groups: the regular Sampler Custom, and the Clown Shark Sampler from RES4LYF, you can swap between the working groups with the Fast Groups Muter.

Plus several of its optional nodes — Detail Boost, Implicit Steps, Momentum and Sigma Scaling. I left notes from the official RES4LYF example workflow for explanations of what they do, along with some quick notes on my testing of them.

I tried most of the samplers and landed on what feels like the sweet spot for Clown Shark: diag_implicit/pareschi_russo_2s with Beta57. It’s a little slower, but totally worth it IMO.

The Klein node for prompting (in System Settings Subgraph) you can just delete from the workflow or disable if you don't want it, but I find it does help.

Quick notes:

• If you have the DasiwaMinimaxH3_dasiwaREF2VAHybridV1 Darksidewalker pulled you are in for a treat when it comes to up close portraits and nearly everything else except for eyes at a distance where it lacks.

• The Beta57 scheduler by RES4LYF is a must for image quality in this workflow.

• For extra detail, run a SeedVR detail pass with the same resolution as the image (no upscaling).

• I’d keep the LoRA strength light as these were made for video.

• Eyes are a problem when distance is involved and prompt adherence spotty, let's hope the image model fixes it.

• If you use JonXL's Generator LORA with text-to-image, I found a strength setting of .20 @8 steps is best, with reference images you can go higher.

• The model I used to create the images was DaSiWa MiniMax H3 and the CLIP was Huihui-Qwen3-8B-abliterated-v2-FP8-Comfy as these both gave much better results.

• Pretty much everything you can tweak is sitting on the subgraphs. Fair warning: cranking the resolution too high will probably throw errors.

• My pin naming is a bit weird, but that’s on purpose. It helps me keep track of what goes where when I’m unplugging and plugging a bunch of stuff.

• I put the Save Image node on a switch, I only save the ones I actually like. Leave the seed on fixed, generate, then flip the switch and hit run to save (or just right-click the preview and save it manually).

• Reference Image input size has a great deal to do with rendering time, the bigger the image the slower the renders. I set the resize @2.0, if you have a slower card set this lower.

Nodes and models you will need to have installed, the list below is the bare minimum.

custom_nodes:

• ComfyUI-MM-1Frame

• ComfyUI-Easy-Use

• ComfyUI-Flux2Klein-Enhancer

• ComfyUI-KJNodes

• RES4LYF

• rgthree-comfy

## Model Links

diffusion_models:

• minimax_h3_ref2va_pruned_int8_convrot.safetensors

text_encoders:

• qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors

vae (HARD REQUIREMENT):

• minimax_h3_t1_image_vae_step1597.safetensors

vae_approx (TAE):

• taeh3.safetensors (MiniMax H3 TAE)

loras/h3:

• minimax_h3_fl2v_lightx2v_turbo_8step_v1.0_resized_avg_rank_24_bf16.safetensors

• MaxiMin-HHH-R2V-ThisIsFine_LoRA_V0_1.safetensors (Optional)

• Minimax H3 Photorealistic Image Generator LoRA (Optional)

## Highly Recommended

diffusion_models:

• DaSiWa MiniMax H3 (REF2VA Turbo or Non-Turbo)

text_encoders:

• Huihui-Qwen3-8B-abliterated-v2-FP8-Comfy