oceanprotocol · giurgiur99 · Aug 4, 2026 · Aug 4, 2026 · Aug 4, 2026 · Aug 5, 2026
@@ -28,9 +28,11 @@ Templates are validated against `ServiceTemplateSchema`
 | `image` + exactly one of `tag` / `checksum` / `dockerfile` | Image spec (`dockerfile` triggers a build and is gated per daemon by `allowImageBuild`)                          |
 | `exposedPorts`                                             | Container ports forwarded to host ports and returned as service endpoints                                        |
 | `command` / `entrypoint`                                   | Docker CMD / ENTRYPOINT overrides                                                                                |
+| `commandFile`                                               | Path to a script, resolved relative to this directory and inlined as `command[0]` at load time (mutually exclusive with `command`) |
 | `envVars`                                                  | Fixed operator-set env vars (values never returned to callers)                                                   |
 | `userConfigurableEnvVars`                                  | Env vars the consumer supplies via ECIES-encrypted `userData` (optional regex `validation`, `sensitive` UI hint) |
 | `requiredResources` / `recommendedResources`               | Gate/score environment selection (`min` is enforced at `SERVICE_START`)                                          |
+| `workflows`                                                 | Selectable graphs for UI-driven templates; each entry is `id` / `name` / `file` (a path to the graph JSON, resolved relative to this directory and inlined into `graph` at load time) |
 
 ## Templates in this folder
 
@@ -101,6 +103,52 @@ so no `command` override is needed. Bundles ComfyUI-Manager for installing check
 custom nodes from the UI; `HF_TOKEN` / `CIVITAI_TOKEN` are optional user env vars for gated
 downloads. ~10 GB VRAM for SDXL.
 
+### `ltx-video-ugc-product.json` — ComfyUI, LTX-2.3 product video (GPU)
+
+ComfyUI preloaded with LTX-2.3 for one fixed workflow (`workflows/ocean_ugc_product.json`): a
+product photo becomes a vertical 9:16 (720×1280) clip with camera motion and ambient audio,
+capped at the graph's 5 seconds (126 frames at 25 fps — trim longer voiceovers or raise the
+frame count). `entrypoint`/`commandFile` override to a bootstrap script
+(`ltx-video-ugc-bootstrap.sh`, inlined as `command[0]` at load time) that downloads the
+~38 GiB weight set (LTX-2.3 22B fp8 checkpoint, Gemma-3-12B text encoder, two LoRAs, the x2
+spatial upscaler) on first launch. Pick a persistent-storage bucket for this template: weights
+live under the bucket, so only the first launch pays the download, and generated clips are
+written to the bucket root (via ComfyUI's `--output-directory`) so the storage API's
+`listFiles` — which only lists the bucket's top-level files — can see them. Without a bucket
+the weights land inside the container and are lost on stop (in `/tmp`, not the bucket).
+
+The bootstrap runs `main.py` directly from the image's read-only bundle with ComfyUI's own
+`--base-directory` flag, pointed at the bucket, so models/input/user/custom_nodes all live
+there and nothing is written under `/root`. This deliberately bypasses the image's entrypoint,
+which copies its bundle into `/root/ComfyUI` and therefore fails on hosts where the service
+container is not uid 0. Relocating `custom_nodes` also means the image's bundled
+ComfyUI-Manager is not loaded — this template ships one fixed workflow and installs no nodes at
+runtime.
+
+The workflow graph is delivered via userData: `COMFY_WORKFLOW_ID` (the workflow's id, e.g.
+`ocean_ugc_product`) plus `COMFY_WORKFLOW` (the graph JSON, gzipped then base64-encoded). The
+bootstrap writes it into a minimal custom-node pack named after that id —
+`custom_nodes/<id>/example_workflows/<id>.json` — which ComfyUI serves at
+`/api/workflow_templates/<id>/<id>.json`, so `?template=<id>&source=<id>` on the ComfyUI URL loads
+it directly. Naming the pack after the id is what lets the client build that link from the template
+alone, without either side hard-coding a path; `source` must be the module name, since ComfyUI
+resolves `source=all` only against its own core templates. The id therefore becomes a directory
+name, so its validation excludes dots (it is imported as a Python module). The graph is also copied
+into `user/default/workflows/`, where it appears in ComfyUI's Workflows sidebar as an ordinary saved
+workflow — no template-browser resolution or URL parameters needed, and the fallback if the deep
+link does not fire.
+Needs a CUDA GPU with 48 GB+ VRAM (LTX-2.3 22B fp8 + Gemma-3-12B encoder).
+
+### `ltx-video-ugc-testimonial.json` — ComfyUI, LTX-2.3 creator testimonial (GPU)
+
+Same image, bootstrap script, and bucket/weight-download behavior as
+`ltx-video-ugc-product.json` above (see that section for details), but with a different fixed
+workflow (`workflows/ocean_ugc_testimonial.json`, delivered via `COMFY_WORKFLOW_ID=
+ocean_ugc_testimonial` + `COMFY_WORKFLOW`): a creator photo plus a voice clip becomes a
+lip-synced vertical 9:16 (720×1280) testimonial, capped at the same 5 seconds (126 frames at
+25 fps). Generated clips likewise land in the bucket root so the storage API can list them.
+Needs a CUDA GPU with 48 GB+ VRAM (LTX-2.3 22B fp8 + Gemma-3-12B encoder).
+
 ### `automatic1111.json` — Stable Diffusion WebUI (A1111) (GPU)
 
 The classic AUTOMATIC1111 UI (`universonic/stable-diffusion-webui`). The image entrypoint

@@ -0,0 +1,101 @@
+#!/bin/bash
+# The image's own entrypoint copies its ComfyUI bundle into /root/ComfyUI, which fails on
+# hosts where the service container is not uid 0 — bypassed here in favor of running ComfyUI
+# straight from the read-only bundle via --base-directory.
+set -euo pipefail
+
+WF_ID="${COMFY_WORKFLOW_ID:-}"
+# COMFY_WORKFLOW_ID is client-supplied and becomes both a directory and a filename.
+# ServiceStartHandler never loads templates, so userConfigurableEnvVars.validation is NOT enforced
+# node-side — this check is the only thing standing between userData and a path traversal. No dots:
+# the directory becomes a Python module name that ComfyUI imports.
+if [ -n "$WF_ID" ] && ! [[ "$WF_ID" =~ ^[A-Za-z0-9_-]{1,64}$ ]]; then
+  echo "[ocean] invalid COMFY_WORKFLOW_ID" >&2
+  exit 1
+fi
+
+echo "[ocean] $(id) | /data/outputs: $(ls -ld /data/outputs 2>&1 | head -1)"
+
+if [ -d /data/outputs ]; then
+  BASE=/data/outputs/comfy
+  OUTPUT_DIR_ARGS="--output-directory /data/outputs"
+else
+  BASE=/tmp/comfy
+  OUTPUT_DIR_ARGS=""
+  echo "[ocean] no bucket mounted — models download to the container and are lost on stop." \
+    "If you did select a bucket, the node is not applying outputBucketId: check that it runs" \
+    "ocean-node with service bucket-mount support, and that persistentStorage is configured." >&2
+fi
+
+MODELS="$BASE/models"
+mkdir -p "$MODELS/checkpoints" "$MODELS/loras" "$MODELS/text_encoders" \
+  "$MODELS/latent_upscale_models" "$BASE/output" "$BASE/input" "$BASE/temp" "$BASE/user"
+
+# Download to .part then rename: a truncated file in a persistent bucket would be treated as
+# cached by every future launch.
+get() {
+  if [ -f "$2" ]; then
+    echo "[ocean] cached $(basename "$2")"
+    return 0
+  fi
+  echo "[ocean] downloading $(basename "$2")"
+  http_code=$(curl -L --retry 5 --retry-delay 5 -C - -o "$2.part" -w '%{http_code}' "$1")
+  if [ "$http_code" = "416" ] && [ -f "$2.part" ]; then
+    echo "[ocean] $(basename "$2").part already complete"
+  elif [ "$http_code" -lt 200 ] || [ "$http_code" -ge 300 ]; then
+    echo "[ocean] download failed for $(basename "$2") (HTTP $http_code)" >&2
+    return 22
+  fi
+  # HTTP 200 is not proof of a model file: a proxy or HF error page returns a few hundred bytes
+  # with a success status. Without this floor that body gets renamed into place and every later
+  # launch treats it as cached. The smallest real file here is ~300 MB, so 10 MB is unambiguous.
+  size=$(wc -c < "$2.part")
+  if [ "$size" -lt 10485760 ]; then
+    echo "[ocean] $(basename "$2") is only $size bytes — not a model file. First bytes:" >&2
+    head -c 300 "$2.part" >&2 || true
+    echo >&2
+    rm -f "$2.part"
+    return 22
+  fi
+  mv "$2.part" "$2"
+}
+
+HF=https://huggingface.co
+get "$HF/Lightricks/LTX-2.3-fp8/resolve/main/ltx-2.3-22b-dev-fp8.safetensors" \
+  "$MODELS/checkpoints/ltx-2.3-22b-dev-fp8.safetensors"
+get "$HF/Comfy-Org/ltx-2/resolve/main/split_files/text_encoders/gemma_3_12B_it_fp4_mixed.safetensors" \
+  "$MODELS/text_encoders/gemma_3_12B_it_fp4_mixed.safetensors"
+get "$HF/Comfy-Org/ltx-2.3/resolve/main/split_files/loras/ltx_2.3_22b_distilled_1.1_lora_dynamic_fro09_avg_rank_111_bf16.safetensors" \
+  "$MODELS/loras/ltx_2.3_22b_distilled_1.1_lora_dynamic_fro09_avg_rank_111_bf16.safetensors"
+get "$HF/Comfy-Org/ltx-2/resolve/main/split_files/loras/gemma-3-12b-it-abliterated_lora_rank64_bf16.safetensors" \
+  "$MODELS/loras/gemma-3-12b-it-abliterated_lora_rank64_bf16.safetensors"
+get "$HF/Lightricks/LTX-2.3/resolve/main/ltx-2.3-spatial-upscaler-x2-1.1.safetensors" \
+  "$MODELS/latent_upscale_models/ltx-2.3-spatial-upscaler-x2-1.1.safetensors"
+
+if [ -n "$WF_ID" ] && [ -n "${COMFY_WORKFLOW:-}" ]; then
+  # Pack directory and graph filename are both $WF_ID, so the client's deep link
+  # (?template=<id>&source=<id>) resolves without either side hard-coding a name.
+  PACK="$BASE/custom_nodes/$WF_ID"
+  SAVED="$BASE/user/default/workflows"
+  mkdir -p "$PACK/example_workflows" "$SAVED"
+  echo 'NODE_CLASS_MAPPINGS = {}' > "$PACK/__init__.py"
+  # Escrow is already claimed by the time this runs — a corrupt payload must not take the container
+  # down with it. Degrade to no workflow and keep going; the decode's stderr lands in the logs.
+  if printf '%s' "$COMFY_WORKFLOW" | base64 -d | gunzip > "$PACK/example_workflows/$WF_ID.json"; then
+    # Also a saved workflow, so it appears in ComfyUI's sidebar without the template browser.
+    cp "$PACK/example_workflows/$WF_ID.json" "$SAVED/$WF_ID.json"
+    echo "[ocean] installed workflow $WF_ID (template pack + Workflows sidebar)"
+  else
+    echo "[ocean] failed to decode COMFY_WORKFLOW — starting ComfyUI without a workflow" >&2
+    rm -f "$PACK/example_workflows/$WF_ID.json"
+  fi
+fi
+
+export HOME="$BASE"
+export PYTHONPYCACHEPREFIX="$BASE/.cache/pycache"
+export XDG_CACHE_HOME="$BASE/.cache"
+export HF_HOME="$BASE/.cache/huggingface"
+
+echo "[ocean] starting ComfyUI with base directory $BASE"
+exec python3.13 /default-comfyui-bundle/ComfyUI/main.py \
+  --base-directory "$BASE" ${OUTPUT_DIR_ARGS} --listen --port 8188 ${CLI_ARGS:-}
@@ -0,0 +1,60 @@
+{
+  "id": "ltx-video-ugc-product",
+  "name": "ComfyUI — UGC product video (LTX-2.3)",
+  "description": "ComfyUI preloaded with LTX-2.3 for vertical short-form product video: a product photo becomes a 9:16 clip with camera motion and ambient audio, capped at the graph's 5 seconds (126 frames at 25 fps). Select a persistent-storage bucket: it holds ComfyUI's whole base directory, so the 38 GiB of weights download once and are reused, and generated clips land in the bucket root. Needs a CUDA GPU with 48 GB+ VRAM.",
+  "image": "yanwk/comfyui-boot",
+  "tag": "cu130-megapak-pt211-20260803",
+  "exposedPorts": [
+    8188
+  ],
+  "entrypoint": [
+    "/bin/bash",
+    "-c"
+  ],
+  "commandFile": "ltx-video-ugc-bootstrap.sh",
+  "workflows": [
+    {
+      "id": "ocean_ugc_product",
+      "name": "Product → vertical clip",
+      "description": "Upload a product photo, get a 9:16 clip with camera motion and ambient audio.",
+      "file": "workflows/ocean_ugc_product.json"
+    }
+  ],
+  "userConfigurableEnvVars": [
+    {
+      "key": "COMFY_WORKFLOW_ID",
+      "validation": "^[A-Za-z0-9_-]+$"
+    },
+    {
+      "key": "COMFY_WORKFLOW"
+    }
+  ],
+  "requiredResources": [
+    {
+      "id": "cpu",
+      "min": 8,
+      "recommended": 16,
+      "unit": "cores"
+    },
+    {
+      "id": "ram",
+      "min": 32,
+      "recommended": 64,
+      "unit": "GB"
+    },
+    {
+      "id": "disk",
+      "min": 40,
+      "recommended": 80,
+      "unit": "GB"
+    },
+    {
+      "kind": "discrete",
+      "type": "gpu",
+      "min": 1,
+      "recommended": 1,
+      "unit": "count",
+      "description": "CUDA GPU, 48 GB+ VRAM recommended (LTX-2.3 22B fp8 + Gemma-3-12B encoder)"
+    }
+  ]
+}
@@ -0,0 +1,60 @@
+{
+  "id": "ltx-video-ugc-testimonial",
+  "name": "ComfyUI — UGC creator testimonial (LTX-2.3)",
+  "description": "ComfyUI preloaded with LTX-2.3 for vertical short-form UGC video: a creator photo plus a voice clip becomes a lip-synced 9:16 testimonial, capped at the graph's 5 seconds (126 frames at 25 fps). Select a persistent-storage bucket: it holds ComfyUI's whole base directory, so the 38 GiB of weights download once and are reused, and generated clips land in the bucket root. Needs a CUDA GPU with 48 GB+ VRAM.",
+  "image": "yanwk/comfyui-boot",
+  "tag": "cu130-megapak-pt211-20260803",
+  "exposedPorts": [
+    8188
+  ],
+  "entrypoint": [
+    "/bin/bash",
+    "-c"
+  ],
+  "commandFile": "ltx-video-ugc-bootstrap.sh",
+  "workflows": [
+    {
+      "id": "ocean_ugc_testimonial",
+      "name": "Creator testimonial (lip-sync)",
+      "description": "Upload a person photo and a voice clip, get a lip-synced 9:16 testimonial.",
+      "file": "workflows/ocean_ugc_testimonial.json"
+    }
+  ],
+  "userConfigurableEnvVars": [
+    {
+      "key": "COMFY_WORKFLOW_ID",
+      "validation": "^[A-Za-z0-9_-]+$"
+    },
+    {
+      "key": "COMFY_WORKFLOW"
+    }
+  ],
+  "requiredResources": [
+    {
+      "id": "cpu",
+      "min": 8,
+      "recommended": 16,
+      "unit": "cores"
+    },
+    {
+      "id": "ram",
+      "min": 32,
+      "recommended": 64,
+      "unit": "GB"
+    },
+    {
+      "id": "disk",
+      "min": 40,
+      "recommended": 80,
+      "unit": "GB"
+    },
+    {
+      "kind": "discrete",
+      "type": "gpu",
+      "min": 1,
+      "recommended": 1,
+      "unit": "count",
+      "description": "CUDA GPU, 48 GB+ VRAM recommended (LTX-2.3 22B fp8 + Gemma-3-12B encoder)"
+    }
+  ]
+}