(Updated - GGUF) Stop Using Nano Banana | DreamOmni2 ComfyUI Is the BEST FREE Workflow!
Added 2025-10-20 17:23:45 +0000 UTC
Link video: https://youtu.be/Iqc45u6lVHo
In this video, I’ll show you how to install and build the DreamOmni2 workflow inside ComfyUI — a powerful, 100% FREE solution that outperforms both open-source and commercial tools, including GPT-5 and Nano Banana.
Step 1: Prepare environment
Install custom node: https://github.com/HM-RunningHub/ComfyUI_RH_DreamOmni2.git
Step 2: Download models
FLUX.1-Kontext-dev: https://huggingface.co/black-forest-labs/FLUX.1-Kontext-dev
DreamOmni2 Models: https://huggingface.co/xiabs/DreamOmni2/tree/main
DreamOmni2 GGUF models for low VRAM (Special thanks to Leo Garraza!): https://huggingface.co/rafacost/DreamOmni2-7.6B-GGUF/tree/main
Please check the workflow in the attachments.
Comments
How to use GGUF model? Where to place it?
Alexander Conferno
2026-02-08 22:35:46 +0000 UTCYes, you can change that string node to the ComfyUI core string node. I use it because it shows the full prompt in one box, which makes it easier for me to adjust and fine-tune.
SOTAI
2026-01-15 17:52:14 +0000 UTChaving a lot of trouble getting this to work as per the guide. When I run the workflow, I get "not a string" exception. I'm wondering about the KepStringLiteral node as that looks old/obsolete vs comfy's own multlline string node? That may be a red herring.
Neil Myatt
2026-01-15 11:19:26 +0000 UTCThanks for the information, let me look into it.
SOTAI
2026-01-07 07:31:08 +0000 UTCany update video regarding DreamOmni3
sekhar Falnama
2026-01-07 07:11:34 +0000 UTCHi, fortunately, the custom node will automatically quantize the Flux Kontext, so even though the original size is 24GB, it will automatically scale down to support lower VRAM. However, you will need to replace the vlm_model with a GGUF file that is suitable for 12GB VRAM (likely Q4). That said, I can’t be 100% sure at the moment since I don’t have access to my local machine right now. Give it a try and let me know how it goes.
SOTAI
2026-01-07 03:53:52 +0000 UTCI'm on a 3080 with 12 Gb Vram, so am I out of luck with this one? The FLUX.1-Kontext-dev is about 24 Gb, so I was hoping to use Flux1-dev-kontext_fp8_scaled instead. But I see no way of modifying the workflow to that end... Advice?
thomas F
2026-01-07 02:35:06 +0000 UTCYes, that’s expected. The more reference images you use, the harder it becomes to control.
SOTAI
2025-12-27 09:16:31 +0000 UTCIt works great when it works, but I find it quite hard to control. Using only one reference image, I can get it to modify that in a lot of different ways, but getting it to take image 2 and 3 into consideration is very difficult. Sometimes it works, but mostly it does not. Similar experiences?
esletner
2025-12-26 10:32:17 +0000 UTCYeah, thank you for sharing the information about the GGUF version of the DreamOmni2 model. I’m really glad someone created it. I’ve also updated my post to acknowledge your contribution so more people can know about it. Thank you so much.
SOTAI
2025-11-18 08:05:20 +0000 UTCThanks for sharing this awesome DreamOmni2 workflow! With my RTX 5060 Ti 16GB, I was getting OOM errors during VLM load. Fixed it by switching to the GGUF version (DreamOmni2-Vlm-Model-7.6B-Q8_0.gguf) from https://huggingface.co/rafacost/DreamOmni2-7.6B-GGUF. Now it runs smooth at 1024x1024 without crashes. Keep up the great content! 🚀
Leo Garraza
2025-11-15 13:59:55 +0000 UTCThis is happening because you’re missing some checkpoint files. Please make sure you’ve downloaded all the required checkpoints, especially the Flux model. Structure folder: https://github.com/HM-RunningHub/ComfyUI_RH_DreamOmni2?tab=readme-ov-file#step-2-download-required-models
SOTAI
2025-11-10 08:56:21 +0000 UTChi im getting Repo id must use alphanumeric chars, '-', '_', or '.'. The name cannot start or end with '-' or '.' and the maximum length is 96:
sekhar Falnama
2025-11-10 07:26:42 +0000 UTC
