Spark Forge
Trosfy/spark-forge
Generative AI local (imagem, 3D, áudio, vídeo) em DGX Spark via Claude Code
Ver no GitHubSkill Claude Code para executar IA generativa local em NVIDIA DGX Spark (Blackwell). Gera imagens com FLUX.2, modelos 3D com Hunyuan3D via ComfyUI local, áudio com ACE-Step e vídeos com WAN/HunyuanVideo, publicando artefatos em S3/CloudFront com URLs assinadas.
Instalação
npx convez-ai Trosfy/spark-forge1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Geração de imagens de alta qualidade (FLUX.2, Z-Image Turbo) localmente sem latência de API
- Conversão image-to-3D (Hunyuan3D) com decimação e otimização de mesh para assets 3D
- Síntese de áudio/música a partir de tags e letras com ACE-Step 1.5
- Produção de vídeos a partir de imagens estáticas ou sequências com WAN 2.2 e HunyuanVideo
Veredito

Implementação técnica sólida para orquestração de modelos generativos em hardware especializado (DGX Spark). Requere infraestrutura NVIDIA específica, conhecimento profundo de ComfyUI e familiares com Terraform; não é para uso casual, mas oferece controle total e latência zero em ambientes enterprise.
README
A tradução para o português deste README ainda está na fila.
spark-forge
Local generative-AI skills for the NVIDIA DGX Spark (GB10 / Blackwell), packaged as a Claude Code plugin. Generate images with FLUX.2 and 3D models with Hunyuan3D — on local weights, through a local ComfyUI — and publish results to S3.
This repo is both a Claude Code plugin marketplace (the forge plugin) and the
Spark setup to stand the stack up. The code is MIT; the models are not — see
THIRD_PARTY.md (several are non-commercial).
What's inside
| Path | What |
|---|---|
plugins/forge/ | The forge plugin: skills forge:image and forge:model, plus the scripts they run |
infra/ | Terraform (and a no-Terraform fallback) for a private S3 artifact bucket |
setup/ | Spark provisioning — model manifest + Hunyuan3D setup (the delta on top of NVIDIA's base) |
config.example | Copy to ~/.config/forge/config |
Skills
forge — use the stack:
/forge:image— text-to-image with FLUX.2, FLUX.2 Klein, or Z-Image Turbo via ComfyUI./forge:model— image-to-3D (GLB) with Hunyuan3D-2.1; optional decimation to low-poly./forge:audio— text-to-music with ACE-Step 1.5 (tags + lyrics → an MP3 track)./forge:video— image-to-video / first-last-frame with WAN 2.2 (14B + 5B) or HunyuanVideo 1.5 (start [+ end] image → MP4)./forge:publish— push a local file/result off the headless box to S3+CloudFront and return a shareable (signed) download URL.
setup — provision the stack:
/setup:cdn— provision the S3 + CloudFront download infra (Terraform) and wire the config; plans first, applies on confirmation./setup:hunyuan— install Hunyuan3D-2.1 (image-to-3D) into the ComfyUI container for/forge:model.
Models & quantizations
forge:image, forge:audio, and forge:video run off one small registry: a family is a ComfyUI graph (code), a model
is a profile (plugins/forge/models/*.json) of shared params, and each model carries a
quants map — one entry per quantization (NVFP4 / BF16 / FP8 / GGUF) that overrides only
the weights that change. Select one with --quant (or FORGE_QUANT):
python3 plugins/forge/scripts/gen_image.py --list-models
python3 plugins/forge/scripts/gen_image.py --model z-image-turbo --quant fp8 --prompt "..." --seed 7
python3 plugins/forge/scripts/gen_image.py --model flux2-dev --print-graph --prompt x --seed 1 # dry-run, no weights
python3 plugins/forge/scripts/gen_audio.py --model ace-step-1.5 --tags "lofi hip hop, mellow keys" --seed 7
Adding a model is a profile JSON; a new architecture is one graph builder in scripts/
registered in model_registry.FAMILIES. --print-graph builds the graph with no ComfyUI or
weights — the same check CI runs across every model × quant.
Prerequisites
- A DGX Spark (or any box) running ComfyUI with the FLUX.2 nodes in a container,
with
input/andoutput/bind-mounted to the host. - Docker, Claude Code, and the AWS CLI (for publishing). Optional: Terraform (bucket IaC) and Blender (mesh decimation).
- Models downloaded per
setup/models.manifest.
Base OS/driver/container provisioning builds on NVIDIA's official
DGX Spark Playbooks; this repo
documents only the generation-stack delta in setup/.
Install (Claude Code plugins)
/plugin marketplace add Trosfy/spark-forge
/plugin install forge@spark-forge # use: image / model / publish
/plugin install setup@spark-forge # provision: cdn (optional)
Add via Git (as above), not a raw file URL — the marketplace serves the plugin by relative path, which only resolves over Git.
Configure (optional)
With ComfyUI in a container, the host input/output dirs are auto-detected from its mounts and the defaults cover the rest — you can start with no config. Add config only to override a default or to set S3:
sudo install -d /etc/forge && sudo cp config.example /etc/forge/config # system-wide
# or: mkdir -p ~/.config/forge && cp config.example ~/.config/forge/config # per-user
Search order: $FORGE_CONFIG > ~/.config/forge/config > /etc/forge/config;
per key, environment > file > built-in default. See config.example for all keys.
Use
In Claude Code:
/forge:image # "generate an image of a brass astrolabe on dark velvet, seed 1234"
/forge:model # "turn that image into a 3D model"
Or call the scripts directly:
python3 plugins/forge/scripts/gen_image.py --prompt "a brass astrolabe on dark velvet" --seed 7
python3 plugins/forge/scripts/gen_model.py --image <path>.png --decimate 1500
Outputs → S3 ("acting as cloud")
The Spark is VPN-only, so results are pushed to S3.
# one-time: create the bucket (Terraform)
cd infra && cp terraform.tfvars.example terraform.tfvars && $EDITOR terraform.tfvars
terraform init && terraform apply
# ...or without Terraform:
./infra/bootstrap_s3.sh my-bucket us-east-1
# publish a result
plugins/forge/scripts/publish_s3.sh <file> --url
Set FORGE_S3_BUCKET / FORGE_S3_REGION in your config. Credentials come from the
standard AWS credential chain (environment, a named profile, SSO, or an instance role)
— provide them however you normally do.
Optional CloudFront (stable, headless download URLs). Presigned URLs expire (≤12h with temporary/STS credentials). For durable links, enable the optional CDN — the bucket stays private and CloudFront serves it via Origin Access Control:
cd infra && terraform apply -var enable_cloudfront=true
# set FORGE_CDN_DOMAIN to the `cloudfront_domain` output; then `publish_s3.sh <file> --url`
# prints https://<domain>/<key> with no expiry. Anyone can run this in their own AWS account.
For gated downloads, enable signed URLs — bucket and CDN stay private and the script mints short-lived signed URLs (signed with a CloudFront key, so unlike the S3 presign they aren't tied to your STS session):
openssl genrsa -out /etc/forge/cloudfront_private_key.pem 2048
openssl rsa -pubout -in /etc/forge/cloudfront_private_key.pem -out /etc/forge/cloudfront_public_key.pem
cd infra && terraform apply -var enable_cloudfront=true -var cloudfront_signed=true \
-var cloudfront_public_key_path=/etc/forge/cloudfront_public_key.pem
# then set FORGE_CDN_KEY_PAIR_ID (cloudfront_key_pair_id output) + FORGE_CDN_PRIVATE_KEY in /etc/forge/config
Terraform options. State is local by default; for shared/remote state copy
infra/backend.tf.example → infra/backend.tf (S3 backend with native lockfile, no
DynamoDB; Terraform ≥ 1.10). Lifecycle is configurable — expire_days,
noncurrent_expire_days, abort_multipart_days. Custom domain: set domain_name +
route53_zone_name and the ACM cert (us-east-1) + DNS records are created and validated
automatically; or pass acm_certificate_arn to reuse an existing cert (e.g. a wildcard).
public_read / --public stays a simpler alternative for direct S3 URLs.
Roadmap
Direction, not commitments — spark-forge ships FLUX.2 + Hunyuan3D today, and the model
registry and two-plugin layout are built to grow. Targets track current, ComfyUI-supported
models — the ones modern accelerators are being optimized for:
- More image models — Stable Diffusion XL, Qwen-Image, Z-Image Turbo, and FLUX.2 Klein, each a registry family (a ComfyUI graph + profiles), added as it's installed and tested.
- Image editing — Qwen-Image-Edit, as a
forge:editskill. - Audio — music generation via ACE-Step 1.5 / 1.5-XL (
forge:audio). setup:models— scripted model-weight provisioning, once there's a clean per-platform source.- Multi-platform — profiles for non-DGX-Spark accelerators (Apple Silicon, AMD Ryzen AI Max+ / Strix Halo); the ComfyUI engine is already platform-agnostic, so it's mostly profiles + setup docs.
- Hunyuan3D durability — bind-mount the repo + weights out of the container's writable layer.
Licenses
Code: MIT. Models: see THIRD_PARTY.md — FLUX.2-dev and Hunyuan3D-2.1 carry their own terms, some non-commercial. Provided as-is as a personal project: no warranty, no support.