Create AI Film-making website -- 2
Budget / Salary$750–1,500
TypeFreelance project
LocationRemote
Posted2 hours ago
Upgrade my existing platform: https://rustling-motion-magic-labs.base44.app/ I need a production-grade AI filmmaking website, not a demo or basic prompt-to-video generator. The system must generate high-quality AI films from 5 seconds up to 2 minutes (120 seconds) and support multiple productions/jobs at the same time.
CORE WORKFLOW: Users must be able to create a project, enter a story/script/commercial concept/prompt, create or upload characters and reference images, create locations, generate/edit a screenplay, automatically create scenes/shots, generate storyboards, approve shots, run Production Preflight, submit production, generate video segments using GPU workers, automatically assemble segments, add dialogue/music/SFX/captions/audio, produce the final MP4 and download the completed film. Support 5/10/15/30/45/60/90/120 seconds and custom durations up to 120 seconds.
LONG-VIDEO ARCHITECTURE: DO NOT attempt to generate a 120-second diffusion video in one request. Automatically divide movies into controlled shots/segments. Example: 120-second film → screenplay → scenes → shots → 5–10 second segments → QC → assembly → audio → upscale/post-processing → final master. A 120-second movie could contain approximately 12–24 segments. Users should not manually assemble them.
BASE44 + GPU BACKEND: Base44 should handle users/login, projects, productions, characters, locations, scripts, storyboards, shots, assets, settings, job status, admin and database. Actual AI generation must run on a dedicated GPU backend. Architecture: Base44 Website → ROLLCALL API → Job Queue → GPU Workers → Storage → QC/Assembly → Final Video. The browser must not be responsible for keeping generation alive.
JOB QUEUE: Implement a real queue using Redis/Celery, Redis/RQ, BullMQ or another production solution. Hierarchy: Production ID → Scene ID → Shot ID → Segment ID → Generation Job ID. Statuses: QUEUED, ASSIGNED, RUNNING, UPLOADING, QC, COMPLETED, RETRYING, FAILED, CANCELLED. Every job must save settings, model, prompt, references, seed, GPU, start/end time, retries, errors and output.
MULTIPLE PRODUCTIONS: Mandatory. I need to submit Production A/B/C/D without waiting for one movie to finish. With one GPU, jobs queue. With multiple GPUs, jobs automatically distribute across GPU-01/GPU-02/GPU-03/GPU-04. Additional GPUs must be addable without rebuilding the website.
GPU WORKERS: Do not hard-code one provider. Support architecture for RunPod, Vast.ai, Lambda, CoreWeave, AWS and local/dedicated servers. Workers register capabilities, GPU model, VRAM, installed models, status and heartbeat. States: ONLINE/BUSY/DRAINING/OFFLINE/ERROR. If a worker crashes, unfinished jobs return to the queue. Completed segments must survive.
VIDEO MODELS: Start with Wan 2.2 or the best appropriate self-hosted model. Support text-to-video, image-to-video and reference-image generation. Build a Model Registry so additional models/APIs can be added without rebuilding the platform. Do not hard-code the database around one model.
CHARACTER IDENTITY LOCK: Character consistency is critical. Each character needs ONE canonical Character ID. Example CAMBELL = CHAR_CAMBELL_001. Every screenplay, scene, shot, storyboard and production must reference this ID rather than typed name variations. Store face/front/left/right/3/4/full-body references, wardrobe, appearance description, approved plates and identity-lock status. Never create duplicate characters. Approved references automatically follow the character into every required generation.
LOCATIONS/WARDROBE/PROPS: Use canonical IDs for locations, wardrobe, props, vehicles and sets. Location records store references, lighting, room/set characteristics, important objects and continuity information. References automatically follow relevant shots.
STORYBOARD + SHOTS: Create storyboard frames before expensive video generation. Allow Approve/Edit/Regenerate/Delete/Reorder. Every shot contains duration, Character IDs, Location ID, action, camera/framing, movement, lighting, wardrobe, props, dialogue, audio, prompt, negative prompt, seed, references and previous/next-shot continuity. Where possible use a previous shot's ending frame as reference for the next shot.
PRODUCTION PREFLIGHT: Before rendering validate characters, character locks, locations, references, storyboards, durations, prompts, GPU worker, model, storage and audio settings. Show READY TO PRODUCE or specific errors explaining what must be fixed. Never silently fail.
QUALITY: Support DRAFT (fast testing), STANDARD (balanced), CINEMA (high quality), MASTER (maximum quality/post-processing). Support 480p draft, 720p generation/preview, 1080p final/upscale and future 4K. Do not falsely call upscaled footage native 4K. Support 16:9, 9:16, 1:1, 4:5 and 2.39:1.
SHOT REGENERATION: If Shot 17 of a 24-shot movie is bad, do NOT regenerate the movie. Allow Regenerate Shot, Change Prompt, Seed, Camera, Motion, Reference or Model. Replace the shot and rebuild the master.
QC + RETRIES: Automatically verify file existence, duration, resolution, FPS, corruption, black/frozen output and where practical AI-check identity, anatomy, unwanted people, wardrobe and location continuity. Failed jobs retry automatically. Attempt 1 FAIL → retry → Attempt 2 → retry → Attempt 3 NEEDS ATTENTION. Never restart an entire production because one shot fails.
CHECKPOINTING: Continuously save production state. If 18/24 shots are completed and a server restarts, continue unfinished work. Never unnecessarily regenerate completed segments.
VIDEO/AUDIO ASSEMBLY: Use FFmpeg or equivalent. Support clip assembly, trimming, transitions, FPS/resolution normalization, dialogue, voice-over, music, SFX, ambience, captions, loudness normalization and final encoding. Architecture must allow future lip-sync/facial performance.
PRODUCTION DASHBOARD: Show production runtime, total shots, completed/rendering/queued/failed shots and overall progress. Show individual shot statuses. Users must be able to pause/cancel production and regenerate individual shots.
ADMIN/GPU DASHBOARD: Show active/queued/failed/completed productions, queue depth, GPU workers, GPU model/VRAM, current jobs, loaded models, worker health, generation time, storage, errors and costs. Track GPU time and estimated cost for each production so I know actual cost for 30/60/90/120-second productions.
API: Create a versioned ROLLCALL API including GET /v1/health, GET /v1/certify, POST /v1/jobs, GET /v1/jobs/{id}, POST /v1/jobs/{id}/cancel, GET /v1/workers, GET /v1/models and POST /v1/worker/heartbeat. Website communicates with the ROLLCALL API, not directly with random Python scripts/ComfyUI workflows. Secure GPU endpoints using server-side API keys/service authentication. Never expose GPU/API secrets in browser code.
STORAGE + DATABASE: Use S3-compatible storage such as S3/R2/B2 for large video/image/audio assets. Database entities should include Users, Projects, Productions, Scenes, Shots, Segments, Characters, CharacterReferences, Locations, Storyboards, Assets, RenderJobs, GPUWorkers, Models, AudioTracks, QCReports, ProductionVersions and Logs. Use resumable/multipart uploads for large media.
NO FAKE FUNCTIONALITY: Every visible Generate, Lock Character, Preflight, Render, Regenerate, Cancel, Pause, Resume, Connect Worker, Test GPU and Download button must perform a real backend action. No placeholders. Do not patch errors individually while leaving broken architecture underneath. Inspect the existing application first and preserve/migrate valid existing data.
ACCEPTANCE TESTS: Before completion demonstrate: (1) real 5-second AI video, (2) assembled 30-second film, (3) 60-second film, (4) complete 120-second film, (5) at least 3 productions submitted simultaneously, (6) queue management, (7) two workers/GPUs processing jobs in parallel, (8) intentionally fail one segment and prove automatic retry, (9) restart worker and prove completed segments survive, (10) regenerate one bad shot without regenerating movie, (11) character references follow every relevant shot, (12) final video/audio synchronization.
DEFINITION OF DONE: This project is NOT complete because a 120-second option appears in a dropdown. It is complete when the platform can reliably go IDEA → SCRIPT → CHARACTERS → STORYBOARD → SHOTS → PREFLIGHT → GPU QUEUE → AI GENERATION → QC → ASSEMBLY → AUDIO → FINAL FILM and return a playable downloadable 30–120 second high-quality AI-generated film. Architecture must support multiple simultaneous productions and multiple GPU workers. Build this as a scalable AI film-production system, not a prototype.
AI Video ← most important
AI App Development
AI Model Development
AI Workflow Automation
Python
Machine Learning (ML)
API Development
Cloud Computing
Full Stack Development
Video Processing
What will be the cost to hourly if we go with your system how much will it cost me per video in 4k or higher quality ive attached a video also know that i have made video witht he site it just keeps dropping or having issues
CORE WORKFLOW: Users must be able to create a project, enter a story/script/commercial concept/prompt, create or upload characters and reference images, create locations, generate/edit a screenplay, automatically create scenes/shots, generate storyboards, approve shots, run Production Preflight, submit production, generate video segments using GPU workers, automatically assemble segments, add dialogue/music/SFX/captions/audio, produce the final MP4 and download the completed film. Support 5/10/15/30/45/60/90/120 seconds and custom durations up to 120 seconds.
LONG-VIDEO ARCHITECTURE: DO NOT attempt to generate a 120-second diffusion video in one request. Automatically divide movies into controlled shots/segments. Example: 120-second film → screenplay → scenes → shots → 5–10 second segments → QC → assembly → audio → upscale/post-processing → final master. A 120-second movie could contain approximately 12–24 segments. Users should not manually assemble them.
BASE44 + GPU BACKEND: Base44 should handle users/login, projects, productions, characters, locations, scripts, storyboards, shots, assets, settings, job status, admin and database. Actual AI generation must run on a dedicated GPU backend. Architecture: Base44 Website → ROLLCALL API → Job Queue → GPU Workers → Storage → QC/Assembly → Final Video. The browser must not be responsible for keeping generation alive.
JOB QUEUE: Implement a real queue using Redis/Celery, Redis/RQ, BullMQ or another production solution. Hierarchy: Production ID → Scene ID → Shot ID → Segment ID → Generation Job ID. Statuses: QUEUED, ASSIGNED, RUNNING, UPLOADING, QC, COMPLETED, RETRYING, FAILED, CANCELLED. Every job must save settings, model, prompt, references, seed, GPU, start/end time, retries, errors and output.
MULTIPLE PRODUCTIONS: Mandatory. I need to submit Production A/B/C/D without waiting for one movie to finish. With one GPU, jobs queue. With multiple GPUs, jobs automatically distribute across GPU-01/GPU-02/GPU-03/GPU-04. Additional GPUs must be addable without rebuilding the website.
GPU WORKERS: Do not hard-code one provider. Support architecture for RunPod, Vast.ai, Lambda, CoreWeave, AWS and local/dedicated servers. Workers register capabilities, GPU model, VRAM, installed models, status and heartbeat. States: ONLINE/BUSY/DRAINING/OFFLINE/ERROR. If a worker crashes, unfinished jobs return to the queue. Completed segments must survive.
VIDEO MODELS: Start with Wan 2.2 or the best appropriate self-hosted model. Support text-to-video, image-to-video and reference-image generation. Build a Model Registry so additional models/APIs can be added without rebuilding the platform. Do not hard-code the database around one model.
CHARACTER IDENTITY LOCK: Character consistency is critical. Each character needs ONE canonical Character ID. Example CAMBELL = CHAR_CAMBELL_001. Every screenplay, scene, shot, storyboard and production must reference this ID rather than typed name variations. Store face/front/left/right/3/4/full-body references, wardrobe, appearance description, approved plates and identity-lock status. Never create duplicate characters. Approved references automatically follow the character into every required generation.
LOCATIONS/WARDROBE/PROPS: Use canonical IDs for locations, wardrobe, props, vehicles and sets. Location records store references, lighting, room/set characteristics, important objects and continuity information. References automatically follow relevant shots.
STORYBOARD + SHOTS: Create storyboard frames before expensive video generation. Allow Approve/Edit/Regenerate/Delete/Reorder. Every shot contains duration, Character IDs, Location ID, action, camera/framing, movement, lighting, wardrobe, props, dialogue, audio, prompt, negative prompt, seed, references and previous/next-shot continuity. Where possible use a previous shot's ending frame as reference for the next shot.
PRODUCTION PREFLIGHT: Before rendering validate characters, character locks, locations, references, storyboards, durations, prompts, GPU worker, model, storage and audio settings. Show READY TO PRODUCE or specific errors explaining what must be fixed. Never silently fail.
QUALITY: Support DRAFT (fast testing), STANDARD (balanced), CINEMA (high quality), MASTER (maximum quality/post-processing). Support 480p draft, 720p generation/preview, 1080p final/upscale and future 4K. Do not falsely call upscaled footage native 4K. Support 16:9, 9:16, 1:1, 4:5 and 2.39:1.
SHOT REGENERATION: If Shot 17 of a 24-shot movie is bad, do NOT regenerate the movie. Allow Regenerate Shot, Change Prompt, Seed, Camera, Motion, Reference or Model. Replace the shot and rebuild the master.
QC + RETRIES: Automatically verify file existence, duration, resolution, FPS, corruption, black/frozen output and where practical AI-check identity, anatomy, unwanted people, wardrobe and location continuity. Failed jobs retry automatically. Attempt 1 FAIL → retry → Attempt 2 → retry → Attempt 3 NEEDS ATTENTION. Never restart an entire production because one shot fails.
CHECKPOINTING: Continuously save production state. If 18/24 shots are completed and a server restarts, continue unfinished work. Never unnecessarily regenerate completed segments.
VIDEO/AUDIO ASSEMBLY: Use FFmpeg or equivalent. Support clip assembly, trimming, transitions, FPS/resolution normalization, dialogue, voice-over, music, SFX, ambience, captions, loudness normalization and final encoding. Architecture must allow future lip-sync/facial performance.
PRODUCTION DASHBOARD: Show production runtime, total shots, completed/rendering/queued/failed shots and overall progress. Show individual shot statuses. Users must be able to pause/cancel production and regenerate individual shots.
ADMIN/GPU DASHBOARD: Show active/queued/failed/completed productions, queue depth, GPU workers, GPU model/VRAM, current jobs, loaded models, worker health, generation time, storage, errors and costs. Track GPU time and estimated cost for each production so I know actual cost for 30/60/90/120-second productions.
API: Create a versioned ROLLCALL API including GET /v1/health, GET /v1/certify, POST /v1/jobs, GET /v1/jobs/{id}, POST /v1/jobs/{id}/cancel, GET /v1/workers, GET /v1/models and POST /v1/worker/heartbeat. Website communicates with the ROLLCALL API, not directly with random Python scripts/ComfyUI workflows. Secure GPU endpoints using server-side API keys/service authentication. Never expose GPU/API secrets in browser code.
STORAGE + DATABASE: Use S3-compatible storage such as S3/R2/B2 for large video/image/audio assets. Database entities should include Users, Projects, Productions, Scenes, Shots, Segments, Characters, CharacterReferences, Locations, Storyboards, Assets, RenderJobs, GPUWorkers, Models, AudioTracks, QCReports, ProductionVersions and Logs. Use resumable/multipart uploads for large media.
NO FAKE FUNCTIONALITY: Every visible Generate, Lock Character, Preflight, Render, Regenerate, Cancel, Pause, Resume, Connect Worker, Test GPU and Download button must perform a real backend action. No placeholders. Do not patch errors individually while leaving broken architecture underneath. Inspect the existing application first and preserve/migrate valid existing data.
ACCEPTANCE TESTS: Before completion demonstrate: (1) real 5-second AI video, (2) assembled 30-second film, (3) 60-second film, (4) complete 120-second film, (5) at least 3 productions submitted simultaneously, (6) queue management, (7) two workers/GPUs processing jobs in parallel, (8) intentionally fail one segment and prove automatic retry, (9) restart worker and prove completed segments survive, (10) regenerate one bad shot without regenerating movie, (11) character references follow every relevant shot, (12) final video/audio synchronization.
DEFINITION OF DONE: This project is NOT complete because a 120-second option appears in a dropdown. It is complete when the platform can reliably go IDEA → SCRIPT → CHARACTERS → STORYBOARD → SHOTS → PREFLIGHT → GPU QUEUE → AI GENERATION → QC → ASSEMBLY → AUDIO → FINAL FILM and return a playable downloadable 30–120 second high-quality AI-generated film. Architecture must support multiple simultaneous productions and multiple GPU workers. Build this as a scalable AI film-production system, not a prototype.
AI Video ← most important
AI App Development
AI Model Development
AI Workflow Automation
Python
Machine Learning (ML)
API Development
Cloud Computing
Full Stack Development
Video Processing
What will be the cost to hourly if we go with your system how much will it cost me per video in 4k or higher quality ive attached a video also know that i have made video witht he site it just keeps dropping or having issues
Apply on Freelancer →
Project sourced from Freelancer.com. Applications happen directly on the original platform — we never collect your data.