~10 min per 10 s clip
Best for everyday clips, drafts and most music-video shots.
mindidlereserved by claude for Attention Files 30s pilot Ep1 · lease ends in 59 min unless renewedOpen ComfyUI ↗
2 leases · updated 02:13:42
What do you want to make?
Video from a prompt — Text prompt + reference images (and optional audio) → a new video clip
Other options
final shots in full 1080p when time doesn't matter~56 min per 10 s clipmind
fast previews and tests~6–9 min per 10 s clipb-dozer
side-by-side comparisons — rarely the right choice~13 min per 10 s clipb-dozer
ideas when you have no reference imagesminutesb-dozer
bringing one picture to lifeminutesb-dozer
rough drafts and multi-shot ideas~70 s per 5 s clipb-dozer
experiments outside the usual pipeline—b-dozer
1080p when speed matters more than perfect timing~24 min per 10 s clipmind
Video to 1080p — Take a clip you already rendered and make it sharper / bigger
~17 min per 10 s clip
Best for turning an approved 640p take into 1080p.
mindidlereserved by claude for Attention Files 30s pilot Ep1 · lease ends in 59 min unless renewedOpen ComfyUI ↗
Other options
only when someone asks for ESRGANREFINE + 1–2 minb-dozermind
the usual way to start a REFINEREFINE + secondsb-dozermind
keeping a 640p take exactly, in 1080p~26 minb-dozermind
Swap people in a video — Replace the people in an existing video with your character
~6–22 min per shot
Best for replacing everyone in a clip with your character, crowds included.
mindidlereserved by claude for Attention Files 30s pilot Ep1 · lease ends in 59 min unless renewedOpen ComfyUI ↗
Other options
faster swaps when hands and set details matter less~6 min per segmentb-dozer
changing just the face (experimental)~5 min per 8 sb-dozer
Images — Text prompt (and optional reference photos) → a still image
~25 s per image
Best for photos, products, posters, text and game art.
mindidlereserved by claude for Attention Files 30s pilot Ep1 · lease ends in 59 min unless renewedOpen ComfyUI ↗
Other options
making refs from a photo for the video workflows~18 s per imageb-dozermind
long scene descriptions and exact layouts~6 s per imageb-dozermind
changing an image using up to 10 referencessecondsb-dozermind
Music — Lyrics + a style description → a full song
~1 min per song
Best for most songs — fast and full length.
mindidlereserved by claude for Attention Files 30s pilot Ep1 · lease ends in 59 min unless renewed
b-dozerin usereserved by claude
Other options
when someone asks for MiniMax Music 3~10–15 min per minute of songb-dozermind
3D models — A picture → a textured, rigged 3D model
minutes
Best for turning a picture into a textured model.
mindidlereserved by claude for Attention Files 30s pilot Ep1 · lease ends in 59 min unless renewed
b-dozernot herenot installed on b-dozer
Other options
getting a model ready to animateminutesmind
rendering or animating scenes—mind
letting an agent build scenes in Blender—mind
Chat / text AI — A local large language model with an OpenAI-style API
instant once loaded
Best for chat, coding help and summaries without the cloud.
b-dozerin usereserved by claude
mindnot herenot installed on mind
Other options
asking questions about picturesinstant once loadedb-dozer
Audio mixing — Mix, master or split finished songs
minutes
Best for finishing songs against a reference track.
b-dozercan startMixMaster isn't running — start it first
mindnot herenot installed on mind
Other options
separating vocals, drums, bass and the restminutesb-dozer
Video from a prompt
Everyday video LARRY6
What it doesThe everyday video workflow: 6-step render from prompt + reference images (+ optional voice/music audio)
Best foreveryday clips, drafts and most music-video shots
How long10 s clip at 1152×640: ~9 min on mind, ~10–12 min on b-dozer
How to open itscript: b-dozer ~/h3-larry-turbo-helper.py or ~/ugc_batch/h3_queue_runner.py (batches); mind ~/h3-singularity-larry-turbo-helper.py
Models it loads
b-dozer
H3 reference-to-video, base (int8)
minimax_h3_ref2va_pruned_int8_convrotLarry Turbo v4 LoRAminimax_h3_turbo_v4_step600_emaH3 text encoder + video/audio VAEsqwen3vl_32b_minimax_h3_nvfp4_awqmind
H3 Singularity v1.3 (int8)
Minimax-h3_Singularity_ref2va_Pruned_v1.3_int8Larry Turbo v4 LoRAminimax_h3_turbo_v4_step600_emaH3 text encoder + video/audio VAEsqwen3vl_32b_minimax_h3_nvfp4_awqRight now
queue b-dozer: would wait behind charswap_mjA_A Watch / queue on b-dozer ↗blocked b-dozer: lease on comfyui held by claude (Attention Files 30s pilot Ep1) Open on b-dozer ↗blocked mind: lease on comfyui held by claude (Attention Files 30s pilot Ep1) Open on mind ↗
Video from a prompt
Sharpest 1080p video LARRY6-1080
What it doesLARRY6 rendered straight at 1920×1088 — the sharpest result
Best forfinal shots in full 1080p when time doesn't matter
How long10 s clip: ~56 min
Watch outmind only (needs its 64 GB RAM); low_vram=1
How to open itscript: mind ~/h3-singularity-larry-turbo-helper.py at 1920×1088
Models it loads
H3 Singularity v1.3 (int8)
Minimax-h3_Singularity_ref2va_Pruned_v1.3_int8Larry Turbo v4 LoRAminimax_h3_turbo_v4_step600_emaRight now
blocked mind: lease on comfyui held by claude (Attention Files 30s pilot Ep1) Open on mind ↗
Video from a prompt
Quick video TURBO4
What it doesFaster 4-step reference-to-video, a little less polished than LARRY6
Best forfast previews and tests
How long10 s clip: ~6–9 min
Watch outb-dozer only (mind has no 4-step LoRA)
How to open itComfyUI workflow: MiniMax-H3-Singularity-Turbo-R2V
Models it loads
H3 Singularity v1.3 (int8)
Minimax-h3_Singularity_ref2va_Pruned_v1.3_int8R2V 4-step turbo LoRAminimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16H3 text encoder + video/audio VAEsqwen3vl_32b_minimax_h3_nvfp4_awqRight now
queue b-dozer: would wait behind charswap_mjA_A Watch / queue on b-dozer ↗
Video from a prompt
Stock video (no speed-up) BASE24
What it doesStock 24-step reference-to-video, no speed-up LoRA (waxy skin, rarely used)
Best forside-by-side comparisons — rarely the right choice
How long10 s clip at 960×544: ~13 min
Watch outb-dozer only
How to open itComfyUI workflow: MiniMax-H3-R2V · script ~/h3-base-24step-helper.py
Models it loads
H3 reference-to-video, base (int8)
minimax_h3_ref2va_pruned_int8_convrotH3 text encoder + video/audio VAEsqwen3vl_32b_minimax_h3_nvfp4_awqRight now
queue b-dozer: would wait behind charswap_mjA_A Watch / queue on b-dozer ↗
Video from a prompt
Video from text only H3-T2V
What it doesText-only video (no reference images), 8-step turbo
Best forideas when you have no reference images
How longminutes
Watch outb-dozer only
How to open itComfyUI workflow: MiniMax-H3-Turbo-T2V (24-step: MiniMax-H3-T2V)
Models it loads
H3 text/first-last-frame-to-video (int8)
minimax_h3_fl2va_pruned_int8_convrotFL2V 8-step turbo LoRAminimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16H3 text encoder + video/audio VAEsqwen3vl_32b_minimax_h3_nvfp4_awqRight now
queue b-dozer: would wait behind charswap_mjA_A Watch / queue on b-dozer ↗
Video from a prompt
Animate a still image H3-I2V
What it doesAnimate a still image (start frame, optional end frame)
Best forbringing one picture to life
How longminutes
Watch outb-dozer only
How to open itComfyUI workflow: MiniMax-H3-I2V
Models it loads
H3 text/first-last-frame-to-video (int8)
minimax_h3_fl2va_pruned_int8_convrotH3 text encoder + video/audio VAEsqwen3vl_32b_minimax_h3_nvfp4_awqRight now
queue b-dozer: would wait behind charswap_mjA_A Watch / queue on b-dozer ↗
Video from a prompt
Very fast video with sound LTX25
What it doesLTX-2.5: very fast video with synced audio, multi-shot cuts in one pass
Best forrough drafts and multi-shot ideas
How long5 s clip at 960×544: ~70 s
Watch outb-dozer only
How to open itscript: b-dozer ~/ltx25_gen.py
Models it loads
LTX-2.5 22B distilled (int8)
ltx-2.5-22b-distilled-transformer-comfy-int8-convrotGemma 4 12B text encodergemma4-12b-with-proj-ltx-2.5-comfy-int8-convrotRight now
queue b-dozer: would wait behind charswap_mjA_A Watch / queue on b-dozer ↗
Video from a prompt
Wan 2.2 video WAN22
What it doesWan 2.2 text/image-to-video (installed, not part of the usual pipeline)
Best forexperiments outside the usual pipeline
How long—
Watch outb-dozer only; no saved workflow — use the ComfyUI template browser (Wan 2.2)
How to open itComfyUI template browser → Wan 2.2
Models it loads
Wan 2.2 14B high+low noise (fp8)
wan2.2_t2v/i2v_*_14B_fp8_scaledLightX2V 4-step LoRAswan2.2_*_lightx2v_4steps_loraUMT5-XXL text encoderumt5_xxl_fp8_e4m3fn_scaledRight now
queue b-dozer: would wait behind charswap_mjA_A Watch / queue on b-dozer ↗
Video from a prompt
Fast native 1080p FREEVIDEO
What it doesNative 1080p with a separate command-line tool (8 steps half-res, then 2 at 1080p)
Best for1080p when speed matters more than perfect timing
How long10 s clip at 1920×1088: ~24 min
Watch outmind only; stop ComfyUI first. Motion runs ~1 s late
How to open itcommand line: mind ~/ai-services/FreeVideo
Models it loads
FlashML FreeVideo
VDN-H3Right now
blocked mind: lease on comfyui held by claude (Attention Files 30s pilot Ep1)
Video to 1080p
Upscale to 1080p REFINE
What it doesUpscale any finished clip to 1920×1088 and re-detail it (4 light steps, tiled)
Best forturning an approved 640p take into 1080p
How long10 s clip: ~17 min on mind
How to open itscript: refine_v2.py — b-dozer ~/hotellobby-upscale/, mind ~/mind-refine/
Models it loads
H3 Singularity v1.3 (int8)
Minimax-h3_Singularity_ref2va_Pruned_v1.3_int8R2V 768p 8-step turbo LoRAminimax_h3_ref2v_turbo_8step_v1.0_768p_comfyui_bf16MMH3 Ultimate Upscale (tiled 2×3)custom nodeRight now
queue b-dozer: would wait behind charswap_mjA_A Watch / queue on b-dozer ↗blocked mind: lease on comfyui held by claude (Attention Files 30s pilot Ep1) Open on mind ↗
Video to 1080p
Upscale, crisper start REFINE-ESR
What it doesREFINE, starting from a RealESRGAN ×4 upscale (crisper; only when asked)
Best foronly when someone asks for ESRGAN
How longREFINE + 1–2 min
How to open itscript: upscale_esrgan.py, then refine_v2.py
Models it loads
RealESRGAN ×4
RealESRGAN_x4plus.pththen the REFINE modelsRight now
queue b-dozer: would wait behind charswap_mjA_A Watch / queue on b-dozer ↗blocked mind: lease on comfyui held by claude (Attention Files 30s pilot Ep1) Open on mind ↗
Video to 1080p
Upscale, plain resize start REFINE-LZ
What it doesREFINE, starting from a plain Lanczos resize (softer; the current default start)
Best forthe usual way to start a REFINE
How longREFINE + a few seconds
How to open itscript: refine_v2.py
Models it loads
ffmpeg Lanczos resize
no modelthen the REFINE modelsRight now
queue b-dozer: would wait behind charswap_mjA_A Watch / queue on b-dozer ↗blocked mind: lease on comfyui held by claude (Attention Files 30s pilot Ep1) Open on mind ↗
Video to 1080p
Render, then upscale LARRY6+REFINE
What it doesRender with LARRY6 at 640p, then REFINE it to 1080p
Best forkeeping a 640p take exactly, in 1080p
How long~26 min on mind
Models it loads
LARRY6 modelsthen the REFINE models
Right now
queue b-dozer: would wait behind charswap_mjA_A Watch / queue on b-dozer ↗blocked mind: lease on comfyui held by claude (Attention Files 30s pilot Ep1) Open on mind ↗
Swap people in a video
Character swap SWAP-LORA
What it doesReplace everyone in a video with your character in one pass (most precise; handles crowds)
Best forreplacing everyone in a clip with your character, crowds included
How long~6 min per short shot on mind, ~22 min per 5 s shot on b-dozer
Models it loads
b-dozer
H3 reference-to-video, base (int8)
minimax_h3_ref2va_pruned_int8_convrotCharacter Swap LoRA (akatz-ai)h3_character_swap_pro4500_1000mind
H3 Singularity v1.3 (int8)
Minimax-h3_Singularity_ref2va_Pruned_v1.3_int8Character Swap LoRA (akatz-ai)h3_character_swap_pro4500_1000Right now
queue b-dozer: would wait behind charswap_mjA_A Watch / queue on b-dozer ↗blocked mind: lease on comfyui held by claude (Attention Files 30s pilot Ep1) Open on mind ↗
Swap people in a video
Pose-based swap SWAP-POSE
What it doesSwap by tracing the body pose, then re-rendering (faster, loses hand/set detail)
Best forfaster swaps when hands and set details matter less
How long~6 min per 7–9 s segment
Watch outb-dozer only (the ControlNet is only installed there; 832×480 max on a 3090)
Models it loads
SDPose pose tracesH3 Fun ControlNet union (int8)
minimax_h3_fun_controlnet_union_pruned_int8_convrotH3 Singularity + Larry v4 LoRARight now
queue b-dozer: would wait behind charswap_mjA_A Watch / queue on b-dozer ↗
Swap people in a video
Face-only swap SWAP-FACE
What it doesChange only the face/neck/hands with a masked repaint (experimental)
Best forchanging just the face (experimental)
How long~4–6 min per 8 s chunk
Watch outb-dozer only (inpaint nodes installed there)
Models it loads
LARRY6 models at partial denoiseH3 AV inpaint nodes
custom nodeRight now
queue b-dozer: would wait behind charswap_mjA_A Watch / queue on b-dozer ↗
Images
Everyday images QWEN-IMAGE
What it doesGeneral-purpose images: photoreal, products, posters, text, game/illustration styles
Best forphotos, products, posters, text and game art
How long2048×1152: ~22 s on mind, ~26 s on b-dozer
How to open itComfyUI workflow: Qwen-Image-2.1-Turbo-8step-T2I (both hosts). b-dozer also: Qwen-Image-2.1-Pruna-8step-T2I (painterly styles), -Original-BF16-T2I / -Official-T2I (40-step base), -GGUF-Q8-T2I, -Transparent-RGBA
Models it loads
Qwen-Image-2.1 Turbo
qwen_image_2.1_turbo_int8_convrot (or _bf16)Qwen3-VL 8B text encoderqwen3vl_8b_int8_convrotQwen-Image VAEqwen_image_2.1_vae_bf16Right now
queue b-dozer: would wait behind charswap_mjA_A Watch / queue on b-dozer ↗blocked mind: lease on comfyui held by claude (Attention Files 30s pilot Ep1) Open on mind ↗
Images
Character reference sheets REFS-QWEN
What it doesEdit from reference photos: character sheets (A-pose, close-up) for the video workflows
Best formaking refs from a photo for the video workflows
How long~17–20 s per image
Watch outalways start from the original photo, never a previous result
How to open itComfyUI workflow: Qwen-Image-2.1-Turbo-8step-RefEdit (both hosts). b-dozer also: -Pruna-8step-RefEdit, -Official-ImageEdit
Models it loads
Qwen-Image-2.1 Turbo
qwen_image_2.1_turbo_bf16Qwen3-VL 8B text encoderqwen3vl_8b_int8_convrotQwen-Image VAEqwen_image_2.1_vae_bf16Right now
queue b-dozer: would wait behind charswap_mjA_A Watch / queue on b-dozer ↗blocked mind: lease on comfyui held by claude (Attention Files 30s pilot Ep1) Open on mind ↗
Images
Fast detailed images FLUX3-KLEIN
What it doesFLUX 3 Klein text-to-image: fast, follows long scene descriptions, layout boxes
Best forlong scene descriptions and exact layouts
How long1024×1024, 15 steps: ~6 s
Watch outuse the *distilled* template; the base Klein model and the prompt upsampler are not installed
How to open itComfyUI template browser → image_flux3_klein_distilled_t2i
Models it loads
FLUX 3 Image Klein, distilled (int8)
flux3_image_klein_sd_int8_convrotQwen3-VL 4B text encoderqwen3vl_4b_int8_convrotFLUX 2 VAEflux2-vaeRight now
queue b-dozer: would wait behind charswap_mjA_A Watch / queue on b-dozer ↗blocked mind: lease on comfyui held by claude (Attention Files 30s pilot Ep1) Open on mind ↗
Images
Edit with reference pictures FLUX3-KLEIN-EDIT
What it doesFLUX 3 Klein image editing with up to 10 reference images
Best forchanging an image using up to 10 references
How longseconds per image
How to open itComfyUI template browser → image_flux3_klein_distilled_image_edit
Models it loads
FLUX 3 Image Klein, distilled (int8)
flux3_image_klein_sd_int8_convrotQwen3-VL 4B text encoderqwen3vl_4b_int8_convrotFLUX 2 VAEflux2-vaeRight now
queue b-dozer: would wait behind charswap_mjA_A Watch / queue on b-dozer ↗blocked mind: lease on comfyui held by claude (Attention Files 30s pilot Ep1) Open on mind ↗
Music
Music 3 song MUSIC3
What it doesMiniMax Music 3: song with vocals from lyrics + caption (use when someone asks for Music 3)
Best forwhen someone asks for MiniMax Music 3
How long1-min song: ~10–15 min
How to open itComfyUI workflow: MiniMax-Music3-T2M (b-dozer) · script: ~/music3-helper.py (mind)
Models it loads
Music 3 DiT (int8)
minimax_music3_dit_int8_convrotMusic 3 text encoder (int8)minimax_music3_text_encoder_pruned_int8_convrotMusic 3 audio decoderminimax_music3_davRight now
queue b-dozer: would wait behind charswap_mjA_A Watch / queue on b-dozer ↗blocked mind: lease on comfyui held by claude (Attention Files 30s pilot Ep1) Open on mind ↗
Music
Song with vocals YUE2-SONG
What it doesYuE2 song with vocals from lyrics + style; the default music engine
Best formost songs — fast and full length
How long2-min song: ~1 min
How to open itcommand line: ~/yue2 (gpu-music skill)
Models it loads
YuE2-3B
Right now
blocked mind: lease on comfyui held by claude (Attention Files 30s pilot Ep1)blocked b-dozer: lease on comfyui held by claude (Attention Files 30s pilot Ep1)
3D models
Picture to 3D model TRELLIS2-3D
What it doesPicture → 3D mesh with texture (3DGenStudio)
Best forturning a picture into a textured model
How longminutes per model
How to open it3DGenStudio (gpu-3d-models skill)
Models it loads
Trellis 2
Right now
blocked mind: lease on comfyui held by claude (Attention Files 30s pilot Ep1)
3D models
Add a skeleton SKINTOKENS-RIG
What it doesAdd a skeleton to a 3D model so it can be animated
Best forgetting a model ready to animate
How longminutes per model; holds ~19 GB while up
How to open it3DGenStudio rigger (gpu-3d-models skill)
Models it loads
SkinTokens / TokenRig
Right now
mind: 3DGenStudio rigger ready
3D models
Blender render BLENDER-RENDER
What it doesRender or animate scenes in Blender
Best forrendering or animating scenes
How long—
Models it loads
Blender (Cycles / EEVEE)
Right now
free to start mind: start Blender
3D models
AI-driven Blender BLENDER-MCP
What it doesLet an AI agent drive Blender over MCP (port 9876)
Best forletting an agent build scenes in Blender
How long—
Models it loads
Blender MCP bridge
Right now
free to start mind: start Blender
Chat / text AI
Local chat AI QWEN38-CHAT
What it doesLocal chat / coding / summarising model with 131K context, OpenAI-compatible API on :8090
Best forchat, coding help and summaries without the cloud
How longholds ~20 GB VRAM while up
How to open it~/ai-services/qwen38-ctl.sh start, then http://100.81.237.88:8090/
Models it loads
Qwen3.8-27B (Q4_K_M)
Qwen3.8-27B-Q4_K_M.ggufRight now
blocked b-dozer: lease on comfyui held by claude (Attention Files 30s pilot Ep1)
Chat / text AI
Chat AI with images QWEN38-VISION
What it doesSame model, reading images
Best forasking questions about pictures
How longinstant once loaded
Models it loads
Qwen3.8-27B (Q4_K_M)
Qwen3.8-27B-Q4_K_M.ggufVision projectormmproj-F16.ggufRight now
blocked b-dozer: lease on comfyui held by claude (Attention Files 30s pilot Ep1)
Audio mixing
Mix & master MIXMASTER
What it doesBatch mix/master songs against a reference and pick winners in a web UI
Best forfinishing songs against a reference track
How longminutes
How to open ithttp://100.81.237.88:8790/
Models it loads
Matchering 2.0
no GPU modelffmpeg loudnormRight now
free to start b-dozer: start MixMaster
~/mixmaster/mixmaster-ctl.sh startAudio mixing
Split into stems STEMS
What it doesSplit a song into vocals / drums / bass / other
Best forseparating vocals, drums, bass and the rest
How longminutes
How to open itMixMaster (demux option) or audio-separator in ~/mixmaster-venv
Models it loads
Demucs
htdemucs_ftRight now
free to start b-dozer: start MixMaster
~/mixmaster/mixmaster-ctl.sh startProcess details
10 GPU processes · per-PID memory, systemd unit, owner, claimb-dozer — GPU processes
| PID | user | service | VRAM | RAM | claim |
|---|---|---|---|---|---|
| 1042893 | bphil | comfyui comfyui.service | 3.1 GB | 12.9 GB | claude |
| 2739541 | ezglove | ezglove-livestream | 1.9 GB | 0.1 GB | unclaimed |
| 525763 | ezglove | ezglove-3dgs-comfyui | 0.2 GB | 0.1 GB | idle |
| 4232 | bphil | desktop | 0.1 GB | 0.0 GB | desktop |
| 4450 | bphil | desktop org.gnome.Shell@x11.service | 0.0 GB | 0.1 GB | desktop |
mind — GPU processes
| PID | user | service | VRAM | RAM | claim |
|---|---|---|---|---|---|
| 141392 | bphil | comfyui | 0.4 GB | 5.8 GB | claude |
| 5745 | bphil | desktop org.gnome.Shell@ubuntu.service | 0.3 GB | 0.5 GB | desktop |
| 7279 | bphil | desktop app-dev.lizardbyte.app.Sunshine.service | 0.3 GB | 0.2 GB | desktop |
| 7473 | bphil | desktop | 0.1 GB | 0.4 GB | desktop |
| 6156 | bphil | desktop org.gnome.Shell@ubuntu.service | 0.0 GB | 0.1 GB | desktop |
Before a long job
Take a lease so others see it is yours (and the dashboard names you instead of "unclaimed"):
gpu-lease run comfyui --host mind --purpose "poster set" --client-id $CID -- …