OBSOLETE: Users should migrate to Ars Mentalia

∞/Infinity - ComfyUI Workflow - Low VRAM & Spec 𝌪𝌵𝌡𝌺𝌝 I am pleased to share ∞ Infinity, a ComfyUI text-to-image workflow for legacy computers and low memory machines. It is designed for computers with 2GB-6GB VRAM to run SDXL derivatives such as Segmind SSD-1B and Segmind Vega. However, it is backwards compatible with SD1.5, fully compatible with SDXL, and can make a secure and low-overhead base to build on for better hardware. 04/05/24 Patch : end step = 10k, promote z-index of CLIP-, wired VAE ifnone backwards. No one is perfect...
05/05/24 Patch : I did not wire it wrong, I just did not understand the node at all. -_-; Sorry about that. Tested & all better now.

˶𝞢⤬⫒ⵖsᐼ˶ ⠑⠭⠙⠽⠎⠁

DEPRECATED: Migrate to Ars Mentalia

Its main options:

• Automatic face and hand correction

• Mid-process VRAM management

• Toggle switch panel

• Precision seed and variation control

• Emits a sound when image completes

• Centralised bus (to simplify rewiring)

• Fits in 1920x1080

• Three-dimensional work flow. Switch layers with this sequence:

CTRL A

ALT C

• Bookmarked areas (overall view/prompt/full screen preview) on keyboard keys:

~ 1 2 3 

• No XL refiner step.

Infinity is meant to be a template and an architecture for others to build on. Feel free to adjust it to your personal needs by adding a second KSampler or model loader if you can manage. I cannot load 6GB files yet.

If you are 4GB or less I highly recommend these settings:

calculate_hash // false
control_before_generate // true
seed action // fixed
batch_size // 1
supersample // false

As well as the following process:

[ S T A G E _ 1 ]
  -facedetail & upscale off
  -revise prompt/sampler/seed
  -moderate step/batch/prompt size
[ S T A G E _ 2 ]
  -edit flow beginning at VRAMs
  -facedetail on, upscale off
  -perform inpainting needs
  -moderate guide sizes
[ S T A G E _ 3 ]
  -edit flow beginning at VRAM step
  -facedetail off, upscale on
  -moderate rescale factor
[ C H A N G E _ M O D E L _ V E R ]
-adjust model, vae, clips, vaenamer node

If you still happen to run out of memory...

[ C U D A _ O U T _ O F _ M E M ]
  - vmem debug on, layered under save
  - click gc_collect or unload_model
  - queue, if fail click on again
  - don't empty_cache after sampler

Instructions are included inside the workflow. Additionally, rather than change XL or SD15 model in a single session, save the session, then close and reopen both the server and the browser

My .bat file:

.\python_embeded\python.exe -s ComfyUI\main.py --normalvram --force-fp16 

Detailed options and wirings are located on the lower layer. It was my attempt to make an artful and orderly workflow to fit within small resolution screens. My intent throughout was to minimize accidental clicks and save time navigating the screen, which can be slow to respond on old processors.

With practice, you can navigate instantly and chain the stages in the queue with no need to think about it. I run on a 1050TI with 4GB RAM and this method allowed me to reliably generate art. Most workflows here never ran on my struggling machine, and instead I would lose several minutes rebooting everything.

This is my first workflow, and I have minimized my mistakes as best as I can. I sincerely hope my work meets to your satisfaction. Thank you!!!

If you enjoy my work, please donate!

Node credits

ltdrdata/ComfyUI-Impact-Pack
ltdrdata/ComfyUI-Inspire-Pack
rgthree/rgthree-comfy
pythongosssss/ComfyUI-Custom-Scripts
Suzie1/ComfyUI_Comfyroll_CustomNodes
wallish77/wlsh_nodes
Trung0246/ComfyUI-0246
BlenderNeko/ComfyUI_ADV_CLIP_emb
receyuki/comfyui-prompt-reader-node
drago87/ComfyUI_Dragos_Nodes
kijai/ComfyUI-KJNodes
davask/ComfyUI_MaraScott_Nodes