ComfyUI 0.43 adds Kling 4.0 as a paid video API node

ComfyUI v0.43.0 shipped on October 27, 2026, and most of it is video. The release adds partner nodes for Kling 4.0, PixVerse V6 Max, Mirelo SFX, HeyGen Voice 1.0 and the ClarityAI Crystal video upscaler, plus 2K and 4K output on FLUX 3 Video. A follow-up, v0.43.1, arrived on October 28 with two more partner nodes. The release notes and the changelog list them all.

Partner nodes are not local models. They are paid API calls that ComfyUI sends from your own machine and bills through your Comfy account in Comfy Credits. Comfy's own documentation is blunt about it: "Partner Nodes require credits for API calls to closed-source models, so they do not support free usage." The app itself stays free, and Comfy states that "ComfyUI will always remain fully open-source and free for local users."

This release is the clearest view yet of that split. The nodes most creators will want are the ones they pay for, and the free part of the update is smaller.

What ComfyUI 0.43 adds

The video additions in v0.43.0 run through Comfy's closed-source partners, and each one needs credits:

Kling 4.0. Kuaishou's model arrives as four nodes: text to video, image to video, reference to video and video edit. Comfy's changelog says the four nodes produce clips of up to 30 seconds with stereo audio generated in the same pass. Reference to video and the edit node are the two that separate Kling 4.0 from a plain text-to-video tool: you can steer a generation with reference images or elements, or take footage you already have and change it.

PixVerse V6 Max. A second model option inside PixVerse's existing text-to-video and image-to-video nodes. The PixVerse V6 nodes generate from 1 to 15 seconds with optional native audio, according to Comfy's pricing page, and V6 Max is the higher-rate tier of the two.

Mirelo SFX 2.0. Sound effects from a text prompt, or timed to the action in a video. Mirelo's own site shows the model watching a clip to place footsteps, impacts and ambience, and stripping music that a video model added without being asked. That second job is the useful one on this beat: a generated clip that bleeds in unwanted music no longer has to be re-rendered.

HeyGen Voice 1.0. Instant voice cloning and text to speech, new in this release.

ClarityAI Crystal video upscaler. A node that upscales finished video, per the changelog entry for the Crystal AI node.

FLUX 3 Video at 2K and 4K. The FLUX 3 Video nodes gained 1440p and 2160p output options alongside their existing 720p and 1080p.

VEED. Comfy also lists VEED's Fabric 1.0, which animates a still photo to a speech track, and Lipsync 2.0, which re-syncs a speaker's mouth to new audio. SLOP TV covered those on October 23 when they first appeared in the nightly build. Genuinely new in the v0.43.1 tag is VEED Dubbing, which the changelog says will "Translate the speech in a video into another language, cloning each speaker's voice and re-syncing their lips to the translation." Comfy also added Patina texture nodes, which generate a tiling PBR material with basecolor, normal, roughness, metalness and height maps from one latent.

Which of these run on your own GPU

Fewer than the list suggests. Every node above is a partner node, so none of it runs on your hardware and none of it downloads weights. VEED's Comfy page spells out the rule for its own nodes and, by extension, the category: "All three are cloud API nodes: they need a Comfy account with credits and download no model files locally." The same page notes that "Each node bills per second or per frame of the clip you send."

The local additions are smaller and mostly about setup rather than new generation:

Fibo 2.0. Bria's image model, now supported natively for text to image and image editing. Comfy repackaged the weights on Hugging Face, where the license is plain: "The model weights are © BRIA AI and licensed under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)." You can run it locally. You cannot sell what it makes without a separate deal with Bria.

MOSS-Image edit LoRAs. LoRA loaders now apply to MOSS-Image's edit model, so community edit adapters load without an outside node pack. Comfy's briefing-watcher Fervor made the point that matters here: a MOSS-Image edit adapter that once needed a node pack and a nightly build reached a few hundred people, and the same adapter inside a stable release reaches everyone who presses update.

ACEScg EXR output. Convert Image Color Space and Save Image (Advanced) now accept ACEScg for EXR output, the color space VFX pipelines pass files around in.

Two utility nodes. Branch selects from connected inputs by name with renameable inputs, and Pad Image pads with a solid color or a mirrored copy of the edges, including padding to a fixed multiple.

What the paid nodes cost

Comfy prices every partner node in credits, and its own cloud page fixes the exchange rate by listing an RTX PRO 6000 at "$3.49/hr" or 736.39 credits an hour. That is about 211 credits to the dollar. Comfy's plan credits land in the same place: $35 a month for the Creator plan includes 7,400 credits a month.

Three of the new nodes have published rates on the Partner Node pricing page:

• PixVerse V6 Max bills per second of output. At 720p it is 12.07 credits a second with audio off and 16.09 with audio on; at 1080p it is 24.14 and 30.84. Comfy flags those as introductory and lists the higher rates that start November 3, 2026.
• Clarity AI Crystal Video Upscale bills 30.17 credits per output megapixel per second. Comfy works the arithmetic in its own footnote: at up to 30 fps, a 1080p output costs about 62.56 credits a second, and the node's 9-megapixel ceiling about 271.56.
• HeyGen Voice 1.0 Text to Speech bills 9.05 credits per 1,000 characters at the standard rate, 2.26 for the turbo model. Both are discounted through October 31, 2026, to about 4.53 and 1.21. The Instant Clone node is free.

The catch is the node with the biggest name. On the current Partner Node pricing page, Comfy has not published a Kling 4.0 credit rate: its Kling section tops out at the o1, 3.0 and 2.6 models. So the cost of the release's headline node is, for now, unknowable from Comfy's own numbers. Thereel Stack, citing a creator's per-clip figures, put a Kling 4.0 shot at about $2.10 against $1.38 for a Wan 3.0 shot, but that is one creator's measuring, not a published rate.

How one graph now carries a clip end to end

Put the paid nodes in a row and the shape of the release becomes clear. One workflow can generate with Kling 4.0 or PixVerse, hand the result to a VEED node to lip-sync or dub it, run Mirelo SFX over it for sound, and finish with the Crystal upscaler, without leaving the graph. Every step bills credits, and the meters differ: PixVerse is per second, Crystal is per megapixel per second, HeyGen is per character. Comfy's credits page warns that one run can briefly leave a negative balance because some partner nodes report cost only after they finish, and while the balance is negative the nodes stay blocked.

Comfy's own advice is to read the price badge on each node before queueing. The changelog adds a matching note for PixVerse: the V6 and V6 Max nodes carry updated price badges, so the cost shown in the graph reflects the current rate.

What 0.43 leaves open

Mirelo SFX and the VEED nodes do not appear in Comfy's credit table, so their per-run cost is only visible in the price badge inside the app. And v0.43.1 is a tag, not an announced release: it carries no GitHub release notes, so the list of what it added comes from the commit history and the changelog, not from a formal post.

The direction is steady. Each of these releases moves another edit step that used to need a second app into a node you pay for by the unit. Comfy keeps the free half free, and the paid half is where the new video work happens.

ComfyUI is a free download from comfy.org, and the new nodes arrive with the in-app update. Sign in under Settings, then User, and check the price badge on each partner node before you queue a batch.