video - https://youtu.be/PRiqTnP3no8
Thanks to the increased resolution, up-scaling takes much longer than generating the initial image... if only there was a way to speed that up a bit?
Well, it looks like Hyper-SD also works great when up-scaling too, and combined with the Euler CFG PP sampler, a latent multiply, some noise injection and a de-noising upscaler (because why not?) you can get some pretty decent images when using just 4 steps in the upscale stage!
Going up to 8 steps does give more detail, but can sometimes be a bit too much. There may even be something like a 2, 4 or 6-step Hyper-SD lora soon too, which would speed things up even more ;)
I've included both the normal & GGUF loader options as it works great with either, and the switch seems to load only the 1 model at a time.
Latent multiply seems to act as a quick and easy built-in node way to reduce (or increase) the colour saturation, while the WAS Nodes noise injection can add some details.
They can also easily be added to any existing up-scaling workflow.
In ComfyUI Manager, the custom nodes installed are:
Impact Pack = ID 2
WAS Node Suite = ID 25
Extra Models for ComfyUI = ID 53
ComfyUI-GGUF = ID 54
Comfyroll Studio = ID 81
Use Everywhere (UE Nodes) = ID 247
As usual, I ran "update all" in manager to make sure everything has the current version.
Attachment info
Flux_Dev_TXT2IMG_HyperSD_MultiLora_NoisyHRF_CFGPP_Compare_v1.json - Loads of nodes. Just used for running a lot of things at once to compare
Flux_Dev_TXT2IMG_HyperSD_MultiLora_NoisyHRF_CFGPP_Basic_v1.json - The basic upscaler, no model loader option
Flux_Dev_TXT2IMG_HyperSD_MultiLora_NoisyHRF_CFGPP_BasicPlus_v1.json - The basic upscaler, including upscale model loader option
As a side note, I'll be away until next week so will answer any questions upon my return!
(update: I'm back! :)
Update - 11/9/2024
Just some minor updates to the new "Flux_Dev_TXT2IMG_HyperSD_MultiLora_NoisyHRF_CFGPP_BasicPlus_LowCFGXtraHiRes_v6.json" file:
Added 2 additional "after detailer" upscale options:
AuraSR - 4x upscale, so 1024 -> x2 = 2048 -> x4 = 8192x8192 in 16 steps
Upscale with model - Pick your own upscale resolution & model
Minimal banding defaults:
Banding is often present on higher resolution images, such as initial images over 1024^2, and if issues are present in the initial image they will be upscaled too
Though slightly slower, the defaults use x2 in 2 sets of 4 steps without the optional upscaling model to minimise banding
Fast group bypasser
As there are now multiple optional groups, there is a "fast group bypasser" node added. This is from the "rgthree" custom node pack - https://github.com/rgthree/rgthree-comfy
Example image showing the workflow results:
Initial Image from the example:
Example 2048x2048 Upscaled Image:
Example 8192x8192 Upscaled Image:
See attachments - ComfyUI_temp_pzpht_00002_.png - its too large ;)
Update 2 - IMG2IMG
Now also attached an IMG2IMG version ;)
Example image showing img2img result:
Update 3 - IMG2IMG + LLM
Now also attached an IMG2IMG + LLM version. Same as above, but LLM powered. See https://www.patreon.com/posts/ai-enhanced-flux-109665789 for more info on LLM party.
Example image showing the workflow results:
Example output:
Other example things for LLMs:
Input an image like this, along with the prompt "do this as a lego character instead"...
... to get a result like this:
It even works with animals too!
Plus text can be included as well!
The really tricky thing is wondering what you're going to re-imagine... ;)