v1.2Version 1.2 wires the LoRA triggers separately for each stage, so you can swap in other LoRAs with different triggers. There's one prompt, which feeds both stages, and two labelled text nodes: "Put Flux 3 lora trigger here" (default "artgainz_colour, ") goes in front of the prompt for the Flux 3 stage, and "Put Flux 2 lora trigger here" (default "ArtGainz. ") goes in front of it for the Flux 2 stage. Flux 2 no longer gets the Flux 3 trigger.Everything else is the same as 1.1: fully latent, one VAE Decode at the end, and preview nodes included but bypassed. Same models, LoRAs and settings as 1.1 (Flux 3 version 1.0 and Flux 2 version 5.1 of my ArtGainz Colour LoRA, each at strength 1.0).The full workflow is embedded in both sample images.v1.1Version 1.1 does everything in one workflow. Flux 3 Klein makes the image and Flux 2 Klein remakes it in the ArtGainz Colour style in the same graph, so you no longer need a separate Flux 3 run or to load an image.It's fully latent. The Flux 3 latent goes straight into Flux 2 (both use flux2-vae, so they share a latent space), and the Flux 2 first pass goes straight into the refine. There's only one VAE Decode, at the very end, so no detail is lost decoding and re-encoding between stages. The Flux 3 hires step now uses a latent upscale (1.5x, bislerp) to stay in latent space. Preview nodes after each stage are included but bypassed; turn them on to watch each step.Flux 3 isn't working very well on its own yet, so this gets you the compositional power of Flux 3 with the look of Flux 2. Both stages use my ArtGainz Colour LoRAs: Flux 3 version 1.0 (file ArtGainzColour_Flux3Klein_2_000002750, strength 1.0, trigger artgainz_colour) and Flux 2 version 5.1 (file ArtGainz_Colour_000002750, strength 1.0, trigger "ArtGainz.").Flux 3 stage: flux3-image-klein-int8-convrot, qwen3vl_4b_int8_convrot text encoder, flux2-vae. lcm, linear_quadratic, 16 steps, CFG 1. Hires (optional): latent upscale 1.5x, 8 steps, denoise 0.45, same seed. The prompt is the Flux 3 prompt upsampler text; I pasted pre-generated text so the 9B upsampler doesn't have to load.Flux 2 stage: flux-2-klein-4b, qwen_3_4b text encoder, flux2-vae. Prompt "ArtGainz. " plus the same upsampled prompt. First pass er_sde, Flux2Scheduler, 20 steps, denoise 0.75, CFG 1 (usable range 0.75 to 0.9; v1.0 used 0.85). Refine: 4 steps, denoise 0.55, same seed.The whole graph takes about 18 s at 768x1024 to 35 s at 1536x1152 on my desktop.Tip: describe everything in detail, including each character's clothing, the time of day and the lighting. The full workflow is embedded in every sample image.v1.0Flux 3 isn't working very well on its own yet, so this workflow lets you get the compositional power of Flux 3 with the look of Flux 2. An img2img workflow that keeps the composition of a Flux 3 Klein image and remakes it in the ArtGainz Colour style with Flux 2 Klein 4B and my ArtGainz Colour LoRA (Flux 2 version 5.1, file ArtGainz_Colour_000002750, strength 1.0). The Flux 3 source images were also made with an ArtGainz colour LoRA (my Flux 3 version 1.0, trigger artgainz_colour). The Flux 3 image goes straight into VAE Encode at its native size, with no reference image. The prompt is 'ArtGainz. ' followed by the source prompt rewritten by the Flux 3 prompt upsampler. I ran the upsampler beforehand, and it isn't a node in this workflow. First pass: er_sde, Flux2Scheduler, 20 steps, denoise 0.85, CFG 1. Usable denoise runs from 0.75 (closer to the source) to 0.9 (more change). Second pass: a 1:1 refine on the same seed at 4 steps, denoise 0.55. Models: flux-2-klein-4b, qwen_3_4b text encoder, flux2-vae. Tip: describe everything in detail, including each character's clothing, the time of day and the lighting. The full workflow is embedded in every sample image.

Restyling Flux 3 Output with Flux 2 Lora

Files with SHA256: 65dee7bb2a76003b055f5c3f12c1fbc84125e2ca5709bc1241e81df8737fef0b