Flux for Low VRAM Users – A Game-Changing Workflow

People often say that the Splashed Mix models match or even surpass Flux when it comes to image quality and aesthetics. That’s an incredible compliment—especially since Splashed Mix models push XL to its limits without extra help.

But what if I could offer my low VRAM user base a way to run Flux too?

Not a fine-tune, not a complicated setup—just a simple workflow that takes a schnellMODE (FLUX.S) model and runs it through a pass of Deep Splashed Something to enhance the image quality. This way, you get the brain of Flux combined with the quality of Splashed.

And the best part?

✅ Runs on just 4GB VRAM
✅ Tested & working – every image in the gallery was generated on a GTX 1650 SUPER (4GB VRAM)
✅ Fast – generates high-resolution images in just over 2 minutes

How to Use the Workflow

1️⃣ Download the Required Models:

• schnellMODE (FLUX.S) Q3 GGUF – the optimized Flux model for low VRAM. Drop it in the ComfyUI/models/unetfolder

• Deep Splashed Something – enhances image quality. Rename it to: deepsplashedsomethingv1.safetensors. Drop it in the ComfyUI/models/checkpoints folder

• Clip_l.safetensors and t5xxl_fp8_e4m3fn.safetensors – place these in your ComfyUI/models/clip folder.

• Grab 4x-Ultrasharp and drop it in the ComfyUI/models/upscale_models folder

2️⃣ Open the Workflow in ComfyUI.

3️⃣ If any nodes are missing, install them using the ComfyUI Manager by clicking the "Missing Nodes" button.

4️⃣ Start prompting! The workflow is designed to be simple and automatic—just enter a prompt and go.

Important Notes

🔹 You don’t need to touch the settings unless you know what you’re doing. The workflow is already optimized for low VRAM users.

🔹 This setup won’t slow you down—it balances quality and performance so you can create stunning images without high-end hardware.

🔹 To keep everything simple the main workflow is a group node but some users have pointed out because the nodes are nested Comfy sometimes throws a hissy about missing nodes. if that happens, copy the nodes from here to get the raw (messy) ungrouped version. Install all the dependecies then the main workflow will play nice.

FAQ

🔹 No LoRA support – This is a pure schnell-base setup. LoRAs won’t work here. The goal is simply to get something out of Flux that Deep Splashed Something can refine into a great image.

🔹 First time sharing a workflow – I’ve tested this, but you might run into issues I haven’t encountered. Drop me a DM or leave a comment, and I’ll fix things ASAP.

🔹 Results may vary – Performance depends on your RAM, CPU, and overall setup. The workflow should work out of the box in most cases, but I can't test every possible configuration.

🔹 Share your results! AI should be for everyone, not just high-end users. Let’s see what we can create on low specs!

🔹 Next step? Getting it running on a calculator. (Joking... for now.)

Special Thanks

All the model makers, hacks, people creating workarounds and bigger brains within the community who do more to push this forward than me. Without all the amazing models and out the box thinking, this would not be impossible.

Huge shoutout to Grumblebutt—one of his images sparked the idea and helped crack the final speed optimisation. Without that missing piece, this wouldn't have been possible!

Have fun :D