<h1 id="about-filmapro-r16cih5he">About FilmaPro</h1>FilmaPro is an experimental project designed to emulate the natural aesthetic of 35mm film, with a bias toward moody photography.Personal NoteThis project is my way of diving deeper into the world of training. It’s an opportunity to learn, experiment, and refine my skills.For anyone who knows what they're doing when it comes to style / concept Loras and has any insights as for how to improve this project please feel free to DM me or leave a comment. I will be leaving my training settings below. If you're interested in this project and would like to support it donating some buzz would be great. <hr /><h3 id="how-to:-rhi0caucy">How To:</h3>Checkout the tags at the end of this page or the example images and have fun! As this is beta some work better than others. Its not perfect.<h3 id="my-settings-jnzqgtlof">My Settings</h3>Checkpoint: Realism_By_Stable_Yogi V4 VAECFG++50 steps0.6 CFGOR DPM ++ 2m5 - 7 CFG20 - 30 stepsRefiner model: Centerfold Version X DMD2 <a target="_blank" rel="ugc" href="https://civarchive.com/models/464549?modelVersionId=1185076">https://civarchive.com/models/464549?modelVersionId=1185076</a>Refine model steps: 40 - 60I use the centerfold dmd2 version as I've noticed it gives me way sharper images. Only downfall with this approach is sometimes if the initial render before the refiner kicks in lacks detail dmd2 has trouble with low detail aspects of an image sometimes.<hr /><h3 id="filmapro-v1-(vibes-+-environments)-lg4m5tf1g">FilmaPro - V1 (Vibes + Environments)</h3>Rating: 7.5/10410 imagesTrained on Realsim_By_Stable_Yogi (v4)Changes: Added vibes + environment tags. Changed training parameters Changed a couple training images Refined existing tags slightly.Overall I'm happy enough with this version to call it v1. Was able to make it up to EPOCH 13 and resolve most of the issues I had with v0.1. Still plenty more to learn and fix but I'll likely wait for some feedback and potentially some constructive criticism before attempting a v1.1 as this stuff gets pricey quick.Overall it seemed like it smoothed out some consistencies from 0.1. For instance direct flash seems more accurate and i get more hits than misses with my intended film style. Let me know what you think. <pre><code>{
 "engine": "kohya",
 "unetLR": 0.00008,
 "clipSkip": 1,
 "loraType": "lora",
 "keepTokens": 0,
 "networkDim": 64,
 "numRepeats": 8,
 "resolution": 1024,
 "lrScheduler": "cosine_with_restarts",
 "minSnrGamma": 4,
 "noiseOffset": 0.1,
 "targetSteps": 8660,
 "enableBucket": true,
 "networkAlpha": 48,
 "optimizerType": "Adafactor",
 "textEncoderLR": 0.00001,
 "maxTrainEpochs": 20,
 "shuffleCaption": true,
 "trainBatchSize": 8,
 "flipAugmentation": false,
 "lrSchedulerNumCycles": 4
}</code></pre><hr /><h3 id="filmapro-v0.1-(beta)-xhg87uud9">FilmaPro - V0.1 (Beta)</h3>Rating: 6/10410 imagesTrained on Realsim_By_Stable_Yogi (v4)I only tagged this version with film aesthetics. Kodak Gold, Soft grain, Harsh Grain, Deep shadows, Faded, Overexposed, underexposed etc. Next version I intend to include more tags.This version certainly does add a decent film aesthetic but from my initial testing I have two areas of concerns. Firstly being it seems some keywords like 'soft grain' ends up being associated with 'faded' and I'm ideally looking to have more granular control over the aesthetics. Secondly, while I trained up to 20 epochs from epoch 6 onward it seems the tags start blending in together making it more difficult to tag the aesthetics appropriately. I'd like to eliminate this issue or at least figure out how to not use so many resources as I'd consider epoch 7 and above mostly a waste. The version included is Epoch 6, I'm going to keep the others to continue testing the higher epochs with different models to see if they may have a use but ultimately I plan to move on to refine this version. There were a couple silly mistakes I've made I'm hoping resolves some issues in the next iteration v0.2.If someone sees something stupid I've done in the training settings please let me know! <pre><code>{
 "engine": "kohya",
 "unetLR": 0.00025,
 "clipSkip": 1,
 "loraType": "lora",
 "keepTokens": 0,
 "networkDim": 64,
 "numRepeats": 6,
 "resolution": 1024,
 "lrScheduler": "cosine_with_restarts",
 "minSnrGamma": 5,
 "noiseOffset": 0.1,
 "targetSteps": 9864,
 "enableBucket": true,
 "networkAlpha": 32,
 "optimizerType": "Adafactor",
 "textEncoderLR": 0.00002,
 "maxTrainEpochs": 20,
 "shuffleCaption": true,
 "trainBatchSize": 5,
 "flipAugmentation": false,
 "lrSchedulerNumCycles": 3
}</code></pre><h3 id="-d9qcsuac2"></h3><hr /><h1 id="filmapro-tags-overview-b91w1jt5n"> FilmaPro Tags Overview</h1>Specific Film Styles<ul><li>Ilford Delta</li><li>Kodak Gold</li><li>Kentmere Pan</li><li>Fujifilm (General)</li><li>Expired Film (General)</li><li>Polaroid (General)</li></ul>Grain Types<ul><li>(grain:0.25)</li><li>(grain:0.5)</li><li>(grain:1)</li></ul>Film Aesthetics<ul><li>Faded</li><li>Light Leak</li><li>Hazy</li><li>Hazy Film</li><li>Soft Focus</li><li>Film Border</li><li>White Border</li></ul>Lighting &amp; Extra Styling<ul><li>Contrast</li><li>High Contrast</li><li>deep shadows</li><li>soft shadows</li><li>soft light</li><li>underexposed</li><li>overexposed</li><li>direct flash</li><li>shadowplay</li><li>backlit</li><li>Shallow depth of field</li></ul>Vibes + Environment<ul><li>Emotive</li><li>Timeless</li><li>Nostalgic</li><li>Expressive</li><li>Dreamy</li><li>Romantic</li><li>Editorial</li><li>Urban</li><li>City</li><li>Beach</li><li>Outdoors</li><li>Indoors</li></ul>

Files with SHA256: 9dc8b91e20a396d80ff35431d22cf692094bf71e2ed2ea6421efb0583db816f2