santa hat
deerdeer nosedeer glow
Sign In

New Emergent Abilities of FLUX.1 Schnell [ image 2 image ]

New Emergent Abilities of FLUX.1 Schnell [ image 2 image ]

UPDATE2: PASSIVE DETAILERS
As previously discussed elsewhere and in my videos - I have trained passive detailers for Flux (Not The True World is my first Passive Detailer. I have trained a couple more, gallery links at bottom!
UPDATE: BOOSTED UPSCALING !
using the same ability explained below, we can exploit this principle in upscaling!
examples are included from V5 onwards: https://civitai.com/models/620294

Flux.1 Schnell is a groundbreaking 4 step base model from Black Forest Labs.

I have been producing my workflow packs to enable artists to hit the ground running with these bleeding edge tools.

In the past with previous diffusion models, when we talk about image 2 image, we use denoise strength to control the output of the image. Generally, 0.5 is considered the middle, which meant that low denoise levels would be typically do very little changes to the output.

As shown in the video below, FLUX.1 schnell exhibits a new ability to boost the prompt guidance using "model shift".

https://civitai.com/models/620294 (workflow pack V3 onwards)

by boosting the model shift we can lower the denoise to as low as 0.1. This allows clean transformation by using higher model shift values. Study the demonstration shown in the video.

In V3 I provided alternate workflows that use only connecting lines (no advanced reference get/set) along with the Flux-Auto-Shift workflow, that used existing prompt enhancers to inject tokens to the Flux Generation. Simple prompts can become denser with only changing a few seeds, while giving control with prefix (subject) and mid (style) prompts entered manually.

If you have not tried out this model yet, take the time to check it out - Don't sleep on Flux.1 !

Would you like to know more?


If you need help setting up FLUX.1 schnell check out the previous videos in this mini-series:
Foda Flux Pack V1
- Multifunction (all-in-one) txt2img/img2img/8Kupscaler & installation details.

Foda Flux Pack V2
- first implementation of Model Shift
- Flux.1 Sampler Garden


Foda Flux Pack V3

Flux Shift Line

  • No Easy use version of v2 Flux Shift (lines only)

Flux Auto Shift

  • Magic prompt + one button prompt nodes to enhance prompts automatically

Flux Auto Shift Line

  • No Easy use version of v3 Flux Auto Shift (lines only)

Foda Flux Pack v4

  • updated to all previous workflows, using low sigma shift and aspect sized dimensions.

  • added new modules for Color correction & RemBG

  • SDXL image Bridge, allows use of SDXL checkpoints,lora,cnets

  • .cube LUTs free downloads added to the bottom of this page

Foda Flux Pack v5

Flux Bridge Kolors

  • Kolors txt2img/img2img with CNET/IPA

  • txt2img/img2img with Flux1

  • BOOSTED Flux Upscaling ! (new feature)

  • all previous modules included

All packs include the pervious versions, new workflows will be added as more capabilities are unlocked

PASSIVE DETAILERS:

I first hit on this with SDXL, however Flux has made this even more impactful, with even more applications for use - to get an idea of how these work, you simply load them at the same time as any other Flux Lora. These work due to the design and curation of the dataset combined with my Tagging/Captioning strategy. Gallery for some of my recent Lora on same seed comparison tables below (~10 images per model, no cherry picking): (big doggo)

https://civitai.com/posts/9095509?returnUrl=%2Fmodels%2F923993%2Fdjz-super-rangers-flux
https://civitai.com/posts/9095689?returnUrl=%2Fmodels%2F930430%2Fdjz-helldivers2-flux
https://civitai.com/posts/9095723?returnUrl=%2Fmodels%2F924046%2Fdjz-super-fashion-flux
https://civitai.com/posts/9095743 (big doggo)
https://civitai.com/posts/9095779?returnUrl=%2Fmodels%2F930524%2Fdjz-star-citizen-flux

https://civitai.com/posts/9095809?returnUrl=%2Fmodels%2F939462%2Fdjz-andromeda-spaceship
https://civitai.com/posts/9095846?returnUrl=%2Fmodels%2F939481%2Fdjz-assassin-zohc
https://civitai.com/posts/9095886?returnUrl=%2Fmodels%2F939500%2Fdjz-assassin-mihq
https://civitai.com/posts/9095903?returnUrl=%2Fmodels%2F939574%2Fdjz-assassin-xuhk
https://civitai.com/posts/9095923?returnUrl=%2Fmodels%2F939522%2Fdjz-cthulu-bishop
https://civitai.com/posts/9095942?returnUrl=%2Fmodels%2F939539%2Fdjz-ultra-future
https://civitai.com/posts/9095967?returnUrl=%2Fmodels%2F939611%2Fdjz-carxjohnson
https://civitai.com/posts/9095994?returnUrl=%2Fmodels%2F939669%2Fdjz-grand-sedan
https://civitai.com/posts/9096272?returnUrl=%2Fmodels%2F944591%2Fdjz-electron-microscopy

The key thing is not to use any trigger, only to load these into the lora stack, then you will benefit from the passively trained data.

Currently this is my List of Passive Detailers:
(however i am testing more that are already released)

- Not the True World Flux
https://civitai.com/models/656031/djz-not-the-true-world-flux
This was the first passive detailer, originally trained with SDXL, but it was less impactful as in Flux - another emergent capability discovery! Far less images were required to get this to work and the effect is far more coherent. NTTW dataset is based on a wide scope of architectural imagery across many different styles. This creates a tendency to favour straight edges and increase the detail depth of any given generation.

- DJZ Abstract Chaos
https://civitai.com/models/930332/djz-abstract-chaos-flux
This is my second passive detailer, this time it was trained with my original dataset created in late 2022, released in Feb 2023 as a SD2.1 Checkpoint. It filed to fully understand what it was being taught however those were early days. Flux has learned this flawlessly and as a result produced and abstract detailed that tends to prefer cursive details, it can also introduce lighting depth and atmosphere to a scene in some situations (spherical falloff)

- More will be added to the list as i complete testing !

15

Comments