Tensor.Art
Create

@m@teur Photography [Flux Dev]

LORA
Reprint

Updated:

V5 final
V3.0
V2.0
V1.0
2.2K

IN NO EVENT SHALL BLACK FOREST LABS, INC. BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH USE OF THIS MODEL.

use it to generate some normal pictures

format of the prompt: "******* photography of <the scene>. Casual, f/<aperture>, <other tags: like low light, slight blur, slight overexposure, warm tones, cluttered background>, jpeg artifacts, on flickr in 2007, 2005 blog, 2007 blog" or you can use any prompt. I feel like the lora doesn't need any trigger words (i maybe wrong though)

i only tested this on flux dev and flux dev q8 gguf. i don't know but when i try other versions like q4 gguf or fp8, the lora doesn't do anything or its only doing minimal changes to the image. i generate directly at 1632x1344 (1088x896 will work also), you can experiment with different resolutions

i used both comfy and forge to test. flux lora training is much better than sdxl. like there is no going back to sdxl or sd3 at this point lol. its my first flux lora, if it doesn't perform well, don't hate me. i will update the lora.

Version Detail

FLUX.1
Recommended Settings (v5-final): Distilled CFG Scale: 2.5 to 4 Sampling method and Schedule type: Heun with BETA or DEIS with DDIM or [Forge] Flux Realistic (Slow) with Beta / DDIM Steps: >=20 (Sometimes I use 20 or 30 or 35 or 40 - You should check at how many steps the image converges - So, please experiment yourself) Resolution: 896x1152 (I made sure it works at this resolution since some of you did not like it when i said generate at high resolutions) - But you have to use hiresfix. Below are the settings I used in my example images Hires. fix: 4x ultrasharp, 0.4 denoise, 10 steps (I'm not the best upscale guy. If you have other upscale methods, you can use it) Checkpoint: flux1-dev-Q8_0.gguf (I use the Q8 one. I don't test this Lora with FP8 or Q4 or other quants) Lora Weight: 0.3-0.5. You have to experiment based on your prompts Positive Prompt: No trigger word needed. But you have to use some photography terms to steer Flux (like photo, photograph etc.). You can start with the below. Its how I tagged the training dataset Full body shot photo of Medium shot photo of Selfie photo of Close-up photo of Or what ever you want (adjust the steps, weight based on your prompt). Detailed prompts still works best Issues: If it happens, please reduce weight Hands Text People in the background Skin texture - its not perfect. I don't like this also. I am trying to find ways to improve it

Project Permissions

Model reprinted from : https://civitai.com/models/652699/amateur-photography-flux-dev

Reprinted models are for communication and learning purposes only, not for commercial use. Original authors can contact us to transfer the models through our Discord channel --- #claim-models.

Comments

Related Posts

Describe the image you want to generate, then press Enter to send.