V2: See version notes. ---&gt;<hr />Originally intended to be a joke placebo image enhancer LoRa which would maybe make images look better (like some models maybe fix faces/hands/image quality). Instead I ended up making a model that actually seemingly straight up improves Flux 1.Dev images across the board (thus HLA / Happy Little Accident).Give it a try and see if you like what you get, and share your generations/thoughts here!Trained on 56 random Flux Pro 1.1 generated images that showcase different details like dynamic lighting, facial details, scenery details, and more.Check out the <a target="_blank" rel="ugc" href="https://civarchive.com/articles/9223/how-to-accidentally-create-a-pretty-good-lora">Article on how this model was made</a>.<hr />Recommended Strength: 0.8 - 1.0Keywords: N/A<hr /><h3 id="image-comparison-quzro99qw">Image Comparison</h3><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/8f3eab25-b752-4cd3-928f-dc88f424d36f/width=525/8f3eab25-b752-4cd3-928f-dc88f424d36f.jpeg" />(Seeds/prompts/image information embedded in model gallery)<hr /><h3 id="original-joke-model-description-tzjrukpsg">Original Joke Model Description</h3>PlaCeBoost: Elevating Image Quality with Perceptual EnhancementTrained off 50 extremely high quality Flux Pro 1.1 images with a focus on details, lighting, depth of field, accuracy, one medium coke, and a large fries, along with the utilization of new Planar-Centralized Boosting techniques (patent pending), images generated using this model should offer a 50%* increase in image quality over base Flux Dev 1.Features:<ul><li>Improved texture of select fruits (tomatoes excluded)</li><li>Enhanced cotton cloth material appearance (polyester blends to come in a later version)</li><li>Increased and decreased foliage</li><li>Better fingers, worse fingernails</li><li>Smarter prompting (user dependent)</li><li>Progressive increase in buzz consumption for users generating multiple images in a row</li><li>Improved quality of shoes (shoelaces to be improved later)</li><li>Smell textures added to certain environments for atmospheric improvements (requires capable PC)</li><li>Vowel rendering improvements (consonants not possible)</li><li>Humans now have the standard 2 nostrils (vs. the FLUX default of 3)</li><li>Beer poured into glass by a proper tap has improved head texture</li><li>Decreased wetness effect of water rendering</li><li>Improved motion blur on blue hedgehogs</li><li>Increased stability</li><li>New hats</li><li>Rainbow rendering now includes UV light for invisible spectrum enthusiasts</li><li>Improved rendering of invisible objects (still invisible, but better)</li><li>Feet are still feet (gross)</li><li>Teeth rendered are brushed using Crest instead of Colgate</li><li>Pizza toppings improved 20% (minimum 18% delivery tip expected)</li></ul>(*Error Margin: 49.5%)

KFT's HLA Image Enhancer [FLUX]

12/02/24: KFT HLA V2Second experiment - nearly halved the size of the original training data set (30) with a focus on keeping only the highest quality images with good lighting, details, etc. This version seems to enhance images quite well, but it doesn't have the same 'creativity' that V1 did. Next test will be revising the smaller image set with more randomness.

KFT_HLA_V2_FLUX.safetensors

1109303_training_data.zip

V2 (30 Image/Detail)

V1 (56 Image/Creative)

Originally intended to be a joke placebo image enhancer LoRa which would maybe make images look better (like some models maybe fix faces/hands/image quality). Instead I ended up making a model that actually seemingly straight up improves Flux <a target="_blank" rel="noopener noreferrer nofollow" href="http://1.Dev">1.Dev</a> images across the board (thus HLA / Happy Little Accident).Give it a try and see if you like what you get, and share your generations/thoughts here!Trained on ~50 random Flux Pro 1.1 generated images that showcase different details like dynamic lighting, facial details, scenery details, and more.Check out the <a target="_blank" rel="ugc" href="https://civitai.com/articles/9223/how-to-accidentally-create-a-pretty-good-lora">Article on how this model was made</a>.Recommended Strength: 0.8 - 1.0Keywords: N/AImage Comparison(Seeds/prompts/image information embedded in model gallery)Original Joke Model DescriptionPlaCeBoost: Elevating Image Quality with Perceptual EnhancementTrained off 50 extremely high quality Flux Pro 1.1 images with a focus on details, lighting, depth of field, accuracy, one medium coke, and a large fries, along with the utilization of new Planar-Centralized Boosting techniques (patent pending), images generated using this model should offer a 50%* increase in image quality over base Flux Dev 1.Features:<ul><li>Improved texture of select fruits (tomatoes excluded)</li><li>Enhanced cotton cloth material appearance (polyester blends to come in a later version)</li><li>Increased and decreased foliage</li><li>Better fingers, worse fingernails</li><li>Smarter prompting (user dependent)</li><li>Progressive increase in buzz consumption for users generating multiple images in a row</li><li>Improved quality of shoes (shoelaces to be improved later)</li><li>Smell textures added to certain environments for atmospheric improvements (requires capable PC)</li><li>Vowel rendering improvements (consonants not possible)</li><li>Humans now have the standard 2 nostrils (vs. the FLUX default of 3)</li><li>Beer poured into glass by a proper tap has improved head texture</li><li>Decreased wetness effect of water rendering</li><li>Improved motion blur on blue hedgehogs</li><li>Increased stability</li><li>New hats</li><li>Rainbow rendering now includes UV light for invisible spectrum enthusiasts</li><li>Improved rendering of invisible objects (still invisible, but better)</li><li>Feet are still feet (gross)</li><li>Teeth rendered are brushed using Crest instead of Colgate</li><li>Pizza toppings improved 20% (minimum 18% delivery tip expected)</li></ul>(*Error Margin: 49.5%)

KFT's HLA Image Enhancer [FLUX] by kftiger

Second experiment - nearly halved the size of the original training data set (30) with a focus on keeping only the highest quality images with good lighting, details, etc. This version seems to enhance images quite well, but it doesn't have the same 'creativity' that V1 did. Next test will be revising the smaller image set with more randomness. 

Files with SHA256: 0a5b7aa5573158fe7a12480bcb4250d4a9ca9ca61f5286a951b696a88d197085