Evaluering av AI-modellerna Flux och Adobe Firefly för matte painting i film
2025 (Swedish)Independent thesis Basic level (degree of Bachelor), 20 credits / 30 HE credits
Student thesis
Abstract [sv]
Idag finns mycket lite akademisk forskning om hur existerande AI-modeller presterar inom deras tänkta område i ett akademisk kontext. Uppdraget var att jämföra två AI-modeller, AdobeFirefly och Flux Fill, inom digital matte painting för film genom inspelning och redigering av fyra olika videoklipp. Videoklippen samlades till ett testkompendie som skickades ut med tillhörande formulär till författarnas samarbetspartner Istudios VFX samt andra individer som är professionellt aktiva inom filmskapande och närliggande fält. Resultaten visade åsikter som var markant jämlika och förtydligar för-och nackdelar hos vardera modell. Deltagarna gav åsikter om videoklippen var redo att brukas i deras nuvarande tillstånd. Majoriteten svarade negativt, men en andel svarade att det krävs vidare justeringar. AI-verktygen behöver vidareutveckling innan de kan bli delar av pipeline utan extern redigering. I dagsläget erbjuder verktyget mycket lite minskning av arbetsbelastningen.
Abstract [en]
Currently, there is very little academic research on the performance of existing AI models in their intended fields. The assignment was to compare the performance of two AI models, AdobeFirefly and Flux Fill, in digital matte painting for film. This was done by recording and editing four different video clips. The video clips were compiled into a test compendium, along with accompanying forms, and sent to the collaborator Istudios VFX, as well as other individuals professionally active in film making and related fields. The results showed markedly similar opinions that clarified the strengths and weaknesses of each model. Participants were asked whether the videos were ready to be used in their current state. The majority answered no, while a percentage answered that further adjustments are needed. These AI tools require further development before they can be integrated into the pipeline without external editing. Currently, the tools offer minimal workload reduction.
Place, publisher, year, edition, pages
2025. , p. 20
Keywords [en]
ComfyUI, Adobe Photoshop, Inpainting, Generative Fill, VFX, Diffusion, Python, Nodes
National Category
Design
Identifiers
URN: urn:nbn:se:bth-28454OAI: oai:DiVA.org:bth-28454DiVA, id: diva2:1987759
External cooperation
Istudios VFX
Subject / course
ME1659 Bachelor Thesis in Media Technology
Educational program
MEGDP Digital Visual Production
Supervisors
Examiners
2025-08-122025-08-072025-09-30Bibliographically approved