Feature Support
This page answers the practical question: which model families support which features today?
Quick Picks
| Need | Best Starting Point |
|---|---|
| LoRA adapters | FLUX.1, SDXL, or Qwen-Image |
| ControlNet | SD 1.5 |
| img2img at 1024 output | FLUX.1 or SDXL |
| broadest feature surface | SD 1.5 or SDXL |
| best prompt-following quality | FLUX.1 or SD 3.5 |
Source Image Workflows
| Family | img2img | Inpainting | Edit / video references |
|---|---|---|---|
| FLUX.1 | Yes | Yes | No |
| SDXL | Yes | Yes | One IP-Adapter image prompt, alongside img2img |
| SD 1.5 | Yes | Yes | One IP-Adapter image prompt, alongside img2img |
| SD 3.5 | Yes | Yes | No |
| Z-Image | Yes | Yes | No |
| Flux.2 Klein | Yes | Yes | No |
| Flux.2 Dev | No | No | Up to 4 ordered native references |
| Wuerstchen v2 | Yes | Yes | No |
| Qwen-Image | Yes | Yes | No |
| Qwen-Image-Edit | No | No | Yes |
| LTX Video | Not yet | Not yet | Not yet |
| LTX-2 | Yes | No | Keyframes |
| Wan Video | N/A | N/A | First/last frame on capable checkpoints |
| MiniMax H3 | N/A | N/A | Required first frame (FL2VA) or ordered refs (Ref2VA) |
| Hunyuan3D | N/A | N/A | Required source image — the only conditioning it has |
Control and Adapters
| Family | Internal family | ControlNet | LoRA |
|---|---|---|---|
| FLUX.1 | flux | No | Yes |
| Flux.2 Klein | flux2 | No | Yes |
| Flux.2 Dev | flux2 | No | No |
| LTX-2 | ltx2 | No | Yes |
| SD 1.5 | sd15 | Yes | Yes |
| SD 3.5 | sd3 | No | Yes |
| SDXL | sdxl | No | Yes |
| Qwen-Image | qwen-image | No | Yes |
| Qwen-Image-Edit | qwen-image-edit | No | Yes |
| Z-Image | z-image | No | Yes |
| Wuerstchen v2 | wuerstchen | No | No |
| LTX Video | ltx-video | No | No |
| Wan Video | wan | No | Yes |
| MiniMax H3 | minimax-h3 | No | No¹ |
| Hunyuan3D | hunyuan3d | No | No |
¹ H3 Turbo model tags contain manifest-pinned, reviewed distillation adapters. They do not enable arbitrary request-time LoRA stacks.
Prompt Conditioning
| Family | Negative Prompts | Scheduler Override |
|---|---|---|
| FLUX.1 | No | No |
| SDXL | Yes | Yes |
| SD 1.5 | Yes | Yes |
| SD 3.5 | Yes | No |
| Z-Image | No | No |
| Flux.2 Klein | No | No |
| Flux.2 Dev | No | No |
| Wuerstchen v2 | Yes | No |
| Qwen-Image | Yes | No |
| Qwen-Image-Edit | Yes | No |
| LTX Video | No | No |
| Wan Video | Yes | Yes¹ |
| MiniMax H3 | No | No |
| Hunyuan3D | N/A² | No |
² Hunyuan3D has no text encoder at all, so neither a prompt nor a negative prompt is read. The source image is the entire conditioning.
¹ Wan's solver set is disjoint from the UNet schedulers: it takes --sample-solver unipc|euler|dpm++, which travels in the same scheduler wire slot. See Wan.
Video Generation
| Family | txt2vid | img2vid | audio2vid | keyframe | retake | lip dub | IC-LoRA | audio track |
|---|---|---|---|---|---|---|---|---|
| LTX Video | Yes | Not yet | No | No | No | No | No | No |
| LTX-2 | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes |
| Wan Video | Yes | Yes | No | First/last | No | No | No | No |
| MiniMax H3 | No¹ | Yes | Ref2VA | First | No | No | No | Required |
| All others | No | No | No | No | No | No | No | No |
¹ The currently advertised compact FL2VA profile requires a first frame. Ref2VA instead requires an ordered set of image, video, or audio references.
LTX Video, LTX-2, and Wan all default to MP4. This lets LTX-2 preserve synchronized audio when requested. Wan renders video only and has no audio path. A build compiled without the mp4 feature falls back to APNG. GIF and APNG remain available for all three families, plus feature-gated WebP. Use --format apng|gif|webp|mp4. Wan keyframing is first/last-frame interpolation only (--image + --last-image); other keyframe layouts are refused at admission. Frame grids are per family: LTX Video and LTX-2 take 8n+1 frame counts (9, 17, 25, 33, ...) with dimensions in multiples of 32; 64 for LTX-2 lip dub, which always renders in two stages and takes its frame count and rate from the reference clip. Wan takes 4n+1 frame counts (49, 53, 81, ...) with dimensions in multiples of 16, except wan22-ti2v-5b, whose 2.2 VAE requires multiples of 32.
The --output extension outranks those family defaults: mold run … -o clip.gif writes a real GIF even where the family would have picked MP4. An extension this binary cannot encode; .mp4 without the mp4 feature, .webp without webp; is refused before any weight is read rather than filled with another container's bytes, and an explicit --format that disagrees with the filename is reported instead of silently overriding it. --output - writes to stdout and claims no extension, so it keeps whatever container the family resolved.
MiniMax H3 is separate from those fallback rules: it always emits MP4 with synchronized generated audio, runs at 24 fps, and accepts 107–345 frames on its 17n+5 grid. It renders one clip per request. See MiniMax H3 for its request and build constraints.
The recommended LTX default today is ltx-video-0.9.6-distilled:bf16. The 0.9.8 family is available, pulls its spatial upscaler asset, and now runs the full multiscale refinement path.
Backend Support
| Family | CUDA | Metal | CPU |
|---|---|---|---|
| FLUX.1 / FLUX.2 | Yes | Yes | Yes (slow) |
| SDXL / SD 1.5 | Yes | Yes | Yes |
| SD 3.5 | Yes | Yes | Yes |
| Z-Image | Yes | Yes | Yes |
| Wuerstchen v2 | Yes | Yes | Yes |
| Qwen-Image | Yes | Yes | Yes |
| Qwen-Image-Edit | Yes | Yes | Yes |
| LTX Video | Yes | Yes | Yes |
| LTX-2 | Yes | Yes | Correctness-only |
| Wan Video | Yes | Yes | Correctness-only |
| MiniMax H3 | SM89 H3 builds | Correctness-only, unqualified | No |
| Hunyuan3D | Correctness-only | Yes | Correctness-only |
Hunyuan3D Metal qualification
Metal is qualified on real weights: hunyuan3d-mini-turbo:fp16 on an M4 Max at octree 192, 256 and 320, compared against ComfyUI on the same checkpoint, image and seed (scripts/capture-hunyuan3d-metal-uat.sh against scripts/capture-hunyuan3d-comfy-metal-reference.sh). The shape path is dense fp16 matmuls and attention with no custom kernel, so CUDA is portable by construction — but "Yes" in this table is a claim about MEASURED performance on real weights, and that campaign has not run on a CUDA card. Correctness-only is the honest answer there until it does.
LTX-2 Metal qualification
LTX-2 / LTX-2.3's Apple Metal path is performance-qualified: BF16 transformer compute, fused attention, streamed FP8 lookup-table widening, and temporal VAE chunks, measured end-to-end on the 19B and 22B distilled FP8 tiers on Apple Silicon. Metal remains slower than a comparable CUDA card; streaming trades speed for fitting the model in unified memory.
Wan Video's Metal path is performance-qualified on Apple Silicon. The measured matrix covers 1.3B BF16 text-to-video at 832x480 and 5B Q8 text-to-video plus 5B FP16 image-to-video at 1280x704, with one cold and three warm 17-frame runs per workload. The path uses family-scoped BF16 and chunked attention; fp8-scaled Wan checkpoints remain refused on Metal.
MiniMax H3's Metal route is shipped but has not completed hardware qualification; the compact stack needs a 64 GB-class Apple Silicon host. H3 generation is otherwise limited to binaries built with the H3 engine, including the Linux SM89 release. Other release targets can still download and verify the registered models and report why generation is unavailable.
Native app surfaces
Desktop and mobile apps use the family capabilities above and the same generation request contract. The table focuses on desktop and iPhone; Android uses the same remote-only Studio surface and native bridge described in the Android guide.
| Area | Desktop | iPhone |
|---|---|---|
| Platforms | macOS, Linux, and Windows (x64 and ARM64) | iPhone and iPad |
| Engine | Built-in local engine (Metal on macOS, CUDA on Linux and on x64 Windows when built with a CUDA toolkit, otherwise CPU; the published Windows nightly is CPU/remote-client only) plus remote hosts | Remote hosts only |
| Create | Capability-driven image/video controls, review-first prepared expansion batches, estimates | Same exact-N, frozen-host review lifecycle and independent siblings, full-screen Advanced sheet, optimized for touch |
| Library | Unified local/remote grid with persisted top-bar thumbnail sizing, host filters, deduplication, batch provenance, a Runs/Prompts History column, and desktop file actions | Merged saved-host library with persisted pinch-to-resize thumbnail sizing, Prints/Collections/Trash scopes, favorites/tag chips, titles via the viewer Info sheet, per-host trash retention on host detail, batch/source provenance, full-screen image/video, swipe navigation, Use as prompt, and Use as source |
| Styles | Ready to use / Browse more with the Still picture · Short clip · 3-D object chips, friendly names over exact ids, kind and explicit 18+ NSFW badges, Get it onto any machine still missing the style (Repair once all have it), download progress | Installed/live union, matching kind/18+ badges and detail, same install-or-repair host picker, pull progress, cancel, load/unload/remove |
| Machines | This device plus remembered/discovered remote hosts, host detail, automatic routing choices | Bonjour, IP/DNS/HTTPS, or Tailscale MagicDNS; explicit generation host and detailed telemetry |
| Queue | Shared multi-host console: live progress, per-job cancel, Pause/Resume, and drag-reorder (queue.can_reorder) | Per-host queue with progress and per-job cancel |
| Settings | A 200px jump nav beside one scrolling page of sixteen always-open sections from the shared schema (Cloud GPUs and Per-style defaults included); rows save on change; ?section= deep links | Six Mold Studio themes with Match phone, remote-host shortcut, version, and TestFlight update channel |
| Sequences | Not authored in the apps — a clip longer than one render is chained and stitched by the machine automatically. Script one with mold run --script or the /api/chain-jobs API. | Same: long clips chain automatically, no scene composer. |
| Updates | Signed Stable/Nightly in-app updater on macOS; Linux and Windows are replaced manually | Internal TestFlight builds after eligible main CI |
See the Desktop App and iPhone App guides for complete workflows.
Notes
- ControlNet is currently available only for SD 1.5.
- LoRA-capable families are
flux,flux2,ltx2,sd15,sd3,sdxl,qwen-image,qwen-image-edit,wan, andz-image. Wuerstchen and LTX Video are not wired for LoRA yet. - Wan adapters cover low-rank pairs and full-weight
.diff/.diff_bdeltas: on bf16 safetensors they merge as the weights are read, on GGUF they apply as a parallel branch at full precision. fp8-scaled Wan checkpoints refuse adapter stacks rather than re-round their weights. - LTX-2 adds stacked LoRAs plus camera-control presets for the published 19B adapters.
--schedulerselects the UNet scheduler for SD 1.5 and SDXL (edm-dpm-pp-2mis Playground v2.5 only). Wan selects its flow solver with--sample-solver unipc|euler|dpm++, which travels in the sameschedulerwire slot; the two sets are mutually exclusive and rejected off-family.- Negative prompts are meaningful for CFG-based families and ignored by FLUX, Z-Image, and Flux.2 Klein. Wan checkpoints were tuned against a specific negative prompt, which mold applies automatically when a request leaves it unset;
/api/modelsadvertises it per model (default_negative_prompt), every surface prefills it, and clearing the field (or--no-negative) sends an explicit empty negative instead. qwen-image-editis a distinct edit family, not a standard img2img mode.- The CLI and API support multiple ordered input images for
qwen-image-edit; the TUI keeps the edit flow to a single source image in v1. qwen-image-editcan use quantized--qwen2-variantlanguage weights while still loading the Qwen2.5-VL vision tower for image conditioning.- LTX-2 now wires
x2spatial upscaling across the family,x1.5spatial upscaling forltx-2.3-*, andx2temporal upscaling in the native runtime. - LTX-2's native CUDA path is validated across text+audio-video, text-to-audio, image-to-video, audio-to-video, keyframe, retake, lip dub, public IC-LoRA, spatial upscale, and temporal upscale workflows.
- LTX-2 renders audio on its own with
--pipeline t2a(pipeline: "t2a"): no video, duration fromframes/fps, and a 16-bit PCM stereowavartifact that lands in the gallery with a rendered waveform tile. - LTX-2's multimodal guider exposes optional per-request overrides for STG scale/blocks, CFG-rescale, cross-modality scale, and the guidance skip stride on the CLI and in web, desktop, iPhone, and TUI Advanced video controls. They apply to the
two-stage,two-stage-hq,keyframe,a2-vid, andt2apipelines; unset fields keep each pipeline's own constants. The TUI uses bounded keyboard cycles for the numeric guidance values and validates comma-separated STG blocks before closing the editor; untouched values remain absent from the request. Its Video accordion also exposes the sharedenable_audiocontract as a checkpoint-aware default/on/off choice, family-gatedspatial_upscale/temporal_upscalenative modes, and the source-freeone-stage,two-stage,two-stage-hq, anddistilledrecipes while Auto leavespipelineabsent. Conditioning-file modes, the audio-onlyt2apipeline, and chain-job administration remain a separate tracked gap. - Completed LTX-2 videos report the runtime-resolved pipeline separately from the requested Auto/explicit choice. Server, CLI, and TUI saves preserve that response in gallery metadata, and web, desktop, iPhone, and TUI Library details show it when present; older and non-LTX prints simply omit the row.
- LTX-2 is performance-qualified on CUDA and Apple Metal (19B/22B distilled FP8 tiers, checkpoint-backed); CPU stays correctness-only.
- Library organization (titles, ♥ favorites, tags, collections, and a trash with per-host retention (
gallery.trash_retention_days)) is stored per host in that host'smold.dband merged across hosts by every client. The web Library exposes it as the Prints | Collections | Trash scope control, the filter-chip row, the print viewer's editable aside, and the selection bar's Add to collection / Tag / Favorite / Trash actions; Settings ▸ Library and Machines ▸ host edit retention. Hosts that do not advertisegallery.organize/gallery.trashshow none of it and keep permanent deletes. See Generating ▸ Browser UI. - Creation-time filing ("File under") lets a print arrive already organized:
GenerateRequestand the chain body carry additivetitle,tags, andcollection, seeded onto the gallery row once, as it is created. The CLI spells itmold run --title/--tag/--collection(with--no-auto-tagand thegenerate.auto_tag_titlepreference); web, desktop, and iPhone Create render a capability-gated File under group, and the TUI keeps it as the last Create ▸ Advanced section. A scripted sequence files the stitched print only, batch and prepared siblings inherit their parent's filing, and a filing the host cannot apply is dropped and reported onx-mold-request-warningrather than failing the render. See Generating ▸ File under.
For model size and VRAM fit, see Models Overview. For usage examples, see Generating Images.