Type | |
Stats | 512 1,990 |
Reviews | (69) |
Published | Sep 6, 2024 |
Base Model | |
Training | Steps: 4,335 Epochs: 15 |
Usage Tips | Clip Skip: 1 Strength: 1.2 |
Trigger Words | 8mmfilm 8mm film |
Hash | AutoV2 1E31E3624B |
Overview
This Lora was trained on 500 images from super 8mm film with 10-20 tags per image. The 8mm Film Lora is trained to capture nuanced analog style attributes like grain, motion blur, and celluloid color. This Lora is quite intense but does a good job of hiding any hallucinations under the analog texture.
Best Use
The sampler & scheduler combination gives a variety of different looks and is important when using this Lora in pushing Flux to stray away from the digital style that it was largely trained on. I have included details below as well as a comparison chart at the end of the image set on this post.
The 8mm Film Lora requires a strong strength to bring out the distinct style of 8mm film and in turn, tends to distort anatomy. The workaround is to pair it with the XLabs realism Lora and to render several iterations at a lower resolution and then go for full resolution and upscaling after a good seed is chosen. All of the images in the dataset were candid and grounded in reality, so if you want to achieve fantasy, sci-fi, or any other distinct genre in conjunction with the 8mm film look it is best to use additional supporting Loras.
I attached a workflow here that is optimized for the Lora that includes upscaling:
https://drive.google.com/file/d/1vA0WlkAOZ-MS6LnUUzeiXAM2iaaWzBWs/view?usp=drive_link
Workflow Settings Summary
8mm Film Lora strength: 1.2
XLabs Realism Lora Strength: 1.2
Flux Guidance: 1.9
Sampler: dpm_fast
Scheduler: ddim_uniform
Max Shift 0.0
Base Shift 4.0
Workflow settings Explained:
-Dimensions: The width and height are set to 1024x768 as optimal dimensions because it outputs a 4:3 image which matches the native 8mm film that all the Lora training data is based on. If you need to go lower to decrease render times I would recommend 512x384. You can change these values, the best results come from setting both values to numbers that are divisible by 16. When you change the values here they are also piped to the upscale group to optimize the settings for the Ultimate SD Upscale node when you choose to use it.
-Lora Strength: ***When using the dpm_fast & ddim_uniform sampler & scheduler Combination (most accurate to real 8mm film) it is best to keep both Lora values at 1.2*** otherwise, for all other sampler & scheduler combinations, set the 8mm Film Lora strength between 2 and 3, the sweet spot is around 2.8, and set the XLabs Realism Lora to 1.0.
-Prompting: This workflow utilizes the Flux dual clip encoders by assigning a separate prompt for each encoder (located in the "playground group"). For the more general clip_L encoder; write prompts that focus on overall key visual elements, such as analog attributes, composition, and stylistic features. Emphasizing textures, lighting, and aesthetic details. For the more ridged T5 encoder; write prompts that emphasize detailed descriptions; capturing the specific physical descriptions of character, environment, spatial relation, narrative, and nuance of the moment.
*Make sure to include the tags 8mm film and 8mmfilm in both prompts. It does not matter whether the tags happen at the start or the end, they just need to be there.
-Max and Base Shift: To extract the maximum texture to emphasize film grain I recommend an unusually low max shift and unusually high base shift. This is because the characteristics of 8mm film are very dramatic; this setting supports the Lora in mirroring that. I prefer max shift at 0.0 and base shift at 4.0.
-Flux Guidance: The sweet spot is 1.9, with 1.5 being the lowest and 2.0 being the highest. Lower values increase the trueness of analog colors and textures but tend to distort anatomy and form. Going above 2.0 encourages the Flux base model to lean on the digital colors and contrast levels it was trained on.
-Scheduler & Sampler:
*dpm_fast & ddim_uniform = slightly blurred and washed out, this combination is the most true to 8mm film with detailed grain and beautifully accurate analog faded colors BUT unlike all of the other scheduler and sampler combinations it needs the following Lora strengths to get results with proper human anatomy:
8mm Film Lora strength: 1.2
XLabs Realisim Lora strength: 1.2\
For all other sampler & scheduler combinations, set the 8mm Film Lora strength between 2 and 3, the sweet spot is around 2.8, and set the XLabs Realism Lora to 1.0.
euler & simple = standard look
huen & normal = medium texture - This is the best combination for simulating the aesthetic of edges of the gate and perforation bleeding into the frame as if the image was over scanned slightly.
dpm_2 & normal = high texture
uni_pc_bh2 & karras = foggy low resolve
dpm_2 & Karras = foggy medium resolve
dpm_fast & sgm_uniform = washed out with motion blur
dpm_fast & beta = washed out and foggy
*Euler & and Beta produce the cleanest images and are a new standard for photorealistic Flux images but are not optimal for this Lora and creating images that feel analog in general.
Upscaling: The images this Lora was trained on were all 2048 max dimensions; the grain and resolution are optimal at 2x with the Ultimate SD Upscaler and give diminishing returns past 3x.
Please share your results, feedback, questions, and buzz me if you are feeling wild :)
-CrunchyBagpipe