Iterative Generative AI Image (0001)

Al geruime tijd ben ik, zoals velen, inclusief fotografen, bezig met het genereren van afbeeldingen/foto’s met AI.
Het geeft vaak inspiratie voor fotoshoots en waar nodig ook hulp bij het bewerken van foto’s.
Vandaag start ik met een nieuw project: Iterative Generative AI.

Het idee hierachter is dat AI, hoewel die prachtige afbeeldingen kan genereren, soms ook totaal de plank mis slaat, ondanks dat de zogenaamde prompt hetzelfde is.

Met de hulpmiddelen van nu start het experiment/project met een door AI willekeurig gegenereerde afbeelding.
Er is geen prompt gegeven.
De app die ik gebruik is Stable Diffusion WebUI Forge en het model waar ik mee werk ik Flux1-dev-bnb-nf4-v2.safetensors.
Voor degenen die ook over dezelfde app beschikken en een computer inclusief grafische kaart hebben die WebUI Forge in combinatie met het Fluxmodel kunnen draaien is het misschien aardig om hieraan mee te doen en te kijken of dezelfde afbeeldingen gegenereerd worden op de eigen computer.

Het doel is om te kijken hoe de gegenereerde afbeeldingen in de loop van de tijd veranderen door het telkens het volgende proces te doorlopen (het iteratieve van het project):
1. De voorgaande afbeelding wordt geüpload naar een willekeurige VLM (Visual Language Model) met de vraag deze zo nauwkeurig mogelijk te beschrijven:
“You’re the best prompt engineer of the world. Describe this photo as detailed as possible in a prompt to be used in the text-to-image app Stable Diffusion WebUI Forge.
Describe the prompt that could recreate this image.
Based on this: give me an outstanding, excellent, precise prompts that lead to a outstanding, extraordinary image in the mentioned text-to-image app.
When done, I’ve a follow-up question for you, so ask me for that.”
2. De door de VLM gegenereerde prompt wordt ingevoerd als prompt in Stable Diffusion WebUI Forge, waarna de afbeelding wordt gegenereerd op grond van de prompt.
3. De follow-up vraag voor de VLM is om één, en slechts één, element uit de prompt te veranderen, toe te voegen of the verwijderen:
” The follow-up question is this: You may change, add or remove a specific element in the prompt, but only one. Then give me your new prompt.”
4. De dan gegenereerde prompt wordt weer ingevoerd in Stable Diffusion WebUI Forge, daar wordt een afbeelding bij gegenereerd en vervolgens wordt die afbeelding weer als “start”afbeelding voorgelegd aan een wilekeurige VLM.

Zal dit uiteindelijk tot een fantastische prompt leiden én tot een prachtige afbeelding?

Voor alle gegenereerde afbeeldingen zijn dit steeds de standaardinstellingen:

Steps 20
Sampler Euler
Schedule type Simple
CFG scale 1
Distilled CFG Scale 3.5
Size 896×1152

Datgene wat zal gaan wijzigen zijn het seednummer en de prompt.

De eerste willekeurig gegenereerde afbeelding, dus de startafbeelding, is deze:

Prompt Geen
Seed 636379820

Voorgelegd aan een VLM komt de volgende afbeelding tot stand:

Prompt A serene, lush tropical forest path with dense green foliage on both sides, scattered dry leaves on a dirt trail, and a misty mountain peak in the background under a cloudy sky. A woman walking away from the camera down the path, wearing a red patterned headscarf and a matching red skirt with white floral designs, a long-sleeve navy blue shirt, and black sandals. She carries a black backpack adorned with small decorative items, including a golden keychain. The scene is vibrant and natural, with tall palm trees and various tropical plants lining the path. The lighting is soft and diffused, evoking a calm, peaceful mood in a remote, rural setting. The perspective is from behind the woman, capturing her mid-step as she walks forward into the lush landscape.
Seed 4152808186

Op de follow-up vraag, antwoordt de VLM met:
Change:
Instead of a cloudy sky, I will change it to a soft golden hour sunlight filtering through the trees, adding warmth and a magical glow to the scene.

En dat levert de volgende afbeelding op:

Prompt A serene, lush tropical forest path with dense green foliage on both sides, scattered dry leaves on a dirt trail, and a misty mountain peak in the background bathed in soft golden hour sunlight filtering through the trees. A woman walking away from the camera down the path, wearing a red patterned headscarf and a matching red skirt with white floral designs, a long-sleeve navy blue shirt, and black sandals. She carries a black backpack adorned with small decorative items, including a golden keychain. The scene is vibrant and natural, with tall palm trees and various tropical plants lining the path. The warm, diffused sunlight creates a calm, peaceful mood in a remote, rural setting. The perspective is from behind the woman, capturing her mid-step as she walks forward into the lush landscape.
Seed 4152808186

Bovenstaande afbeelding zal dus weer als “start” afbeelding geüpload worden naar een willekeurige VLM om de afbeelding zo goed mogelijk te beschrijven.

Regelmatig, maar niet op vaste dagen, zal ik de volgende nieuwe afbeelding van dit project hier op de website plaatsen.

Voor alle berichten met betrekking tot dit project, volg deze link.
Andere AI gegenereerde afbeeldingen uit 2025 en 2024.