Versely
    Workflows

    Remove the background before you upscale

    Cut out the subject first, then upscale the keeper. Upscaling a busy background wastes pixels the I2V pass will animate.

    Versely Team••11 min read

    Cut the subject out first. Then upscale that keeper. Then, if the job is motion, run image-to-video. Upscaling a busy background spends the pass on wallpaper, clutter, and bokeh you are about to throw away. Image-to-video will animate those pixels if you leave them in.

    This page is the order for a still that will become a clip: isolate, enlarge, then move. It is not "what an upscaler is." It is not "when to 4K a locked video." Those are other jobs.

    Busy plates spend the upscale on the room

    An upscaler does not know which object you care about. The published still rows take the whole file. Clarity Crystal Upscaler is 4 credits, 4K listed, image required. SeedVR Upscale is 4 credits, 4K listed, image required. Topaz Upscale Image is 8 credits, 4K listed, image required. None of those cards ship a "subject only" toggle. You feed a still. You get a larger still.

    Clarity's catalog line is honest about what it does with those extra pixels: it regenerates fine texture rather than stretching. That is useful on a product label. It is waste on a kitchen backsplash you will key out. SeedVR's card is a 4K still pass with aspect chips. Topaz's card is the detail-faithful enlargement. All three will faithfully enlarge a messy desk. Faithful is the problem.

    Count the frame, not the feeling. If the bottle occupies a third of the pixels, two thirds of the new texture lands on the room. The upscaler will invent grain in the grout, sharpen a fridge magnet, and crisp a window reflection that is not the SKU. You then either crop that room away (you paid to invent it) or you hand the 4K mess to image-to-video, which treats the room as part of the shot.

    Image-to-video is not a crop tool. The still is frame one. Composition mistakes in the still are permanent. A soft still yields a soft clip. A cluttered still yields a clip that tries to move the clutter. Prompt work will not rescue a plate you should have cleaned.

    The cheap habit is "sharpen it so we can see." Seeing a product does not require 4K wallpaper. Seeing a product requires the product to fill the useful area, at working resolution, with a matte you have already inspected. Pixels come after that.

    If you are generating the still rather than shooting it, make the plate simple on purpose. Text to image is the generate door: one prompt, catalog image models, then edit or send the still into video. Ask for seamless paper, a studio sweep, or a single brand colour behind the subject. A generate that already isolates the object is a cutout you did not have to run. A generate that invents a lifestyle kitchen is a cutout you now owe.

    Cut, inspect, crop, then enlarge

    Do the cheap, reversible work while the file is still small.

    1. Isolate the subject. For a photo on your machine, that is the background remover. It is a free on-device BiRefNet-lite cutout (WebGPU when the browser has it, WASM if not). You drop a PNG, JPEG, WebP, or HEIC. You get a transparent PNG. The image never uploads to Versely. First run caches model assets from Hugging Face in the tab. Product and person shots with a clear subject work best. A busy lifestyle photo with three competing objects will make the segmenter guess. Help it: one subject, contrast against the plate.

    The free tools shelf already pairs this with the upscaler. The free image upscaler lists the remover underneath with the note to cut out before you upscale. That pairing is the product, not a slogan. The upscaler is Anime4K on-device (2x, 3x, or 4x when WebGPU is present, canvas upsample plus sharpen if it is not). Use it as a check on the cutout, not as a way to "preview" a plate you have not isolated.

    Versely's paid photo path is different from the video path, and you should not mix them. remove_background is built for footage. For a still, remove the background from a photo runs as a prompt-driven edit on generate_image_from_image. You attach the photo. You ask for a transparent backdrop, or you describe a replacement plate. You pick an edit-capable model. It is billed as that edit, not as a dedicated one-click eraser. Review the edges before you spend the next credit. Hair, glass, and motion blur are where a general edit snaps toward a helmet. If the free BiRefNet matte is cleaner, keep the free file.

    2. Inspect the matte before you buy pixels. Zoom the cutout at working resolution. Flyaways should not be a cardboard outline. Spectacle rims should not grow a dark crust. A wine glass should not become a bottle-shaped hole. If the alpha is wrong now, a 4K version of it is still wrong, only heavier. Fix the matte, reshoot on paper, or regenerate the still. Do not upscale a helmet and hope.

    3. Crop to the delivery shape at native pixels. The image cropper is on-device. Drag a box, export PNG at the source's own width and height, no upload, no watermark. Do this after the cutout and before the upscale. If the listing is 9:16, crop 9:16 now. If the master is a square pack shot, crop square now. Upscaling the uncropped still and then throwing away half of it is paying for pixels you delete. The cropper exists so the next pass only sees the keep area.

    4. Upscale the isolated keeper. Upscale an image's resolution is the catalog job (upscale_image): an image URL, a model, a larger file of the same picture. Clarity Crystal at 4 credits, SeedVR at 4, Topaz Image at 8. Confirm the sticker before you run it. For a local check with no credits, the free Anime4K tool is enough to decide whether the cutout holds. For a print file or a plate you will animate, use the catalog row that lists the ceiling you actually need. 4K is listed on those three still cards. If delivery is a 1080p Reel, you may not need 4K. Spend the pass when the next step needs the pixels.

    Do not insert an upscale between cutout attempts. Each matte revision should see the logical picture. A twice-enlarged intermediate is a new file the next edit will reinterpret.

    I2V inherits frame one, including clutter

    Once the still is the still, motion is a different family.

    Wan 2.7 Image to Video is 15 credits on the published card, 1080p listed, image required, durations 2s through 15s, audio null. Kling Video V3 Standard Image to Video is 63 credits, image required, audio true, durations 3s through 15s. LTX 2.3 Image to Video Pro is 6 credits, 4K listed, image required, 6s / 8s / 10s, audio null. Those are generate bills. They animate the plate you attach. They do not know you "meant" the bottle and not the curtain.

    That is why the cutout belongs before the upscale, and the upscale belongs before I2V. A working-res cutout lets you see identity. The upscale puts detail on the subject so frame one is not soft. I2V then inherits a clean, sharp opening frame. Reverse any two of those and the expensive row (15, 63, or 6 credits, plus whatever duration you pick) is spent on a plate you would not print.

    Two practical constraints sit between the PNG and the I2V attach:

    Transparent is not always a scene. The free remover exports alpha. Many image-to-video rows want a full RGB plate. A checkerboard composite, or a black fill you did not choose, can become "the room." Composite the upscaled cutout onto a controlled backdrop (seamless white, grey, or a brand colour) unless you have already proven that this model accepts alpha and leaves it alone. The composite is still a still. It is cheaper to retry than a clip.

    The prompt is stage direction. The picture already exists. Describe motion, not the object: a slow product turn, a gentle push-in, steam that does not exist in the plate if you do not want steam invented. Re-describing the bottle in the prompt is how I2V drifts the label to match your wording. The cutout did the isolation. Do not argue with it in the prompt.

    If the brief is a new environment (kitchen, gym bag, beach towel) rather than a turn of this exact frame, that is reference-to-video, not I2V. I2V pins frame one. A cluttered lifestyle photo as a reference also makes the model guess the subject, which is the same isolation problem with a different verb. This page still wants the cutout first. The motion family is a later choice.

    Do not I2V the original messy JPEG "to see if it moves," then plan to key the clip. Bria Video Background Removal is 1 credit on the published card. VEED Video Background Removal is 3 credits, image required, a per-second video job. Those rows exist for footage you already approved. Using them to delete a room you could have deleted as a still is how a 15-credit or 63-credit generate becomes a second bill for work the PNG already did.

    After the clip, 4K video is a different pass

    The stills order stops when the take is the take. If delivery then wants 4K pixels on footage, that is a video upscaler, not another still pass.

    ByteDance upscaler is 1 credit 4K is that job: bytedance-upscaler-video, edit-video, 4K listed, 1 credit on the published card, empty duration list, audio null. Feed a clip you are keeping. Do not feed a JPEG. Do not run it on ten I2V maybes so a client can "see them properly." Wan's I2V card already lists 1080p. Plenty of Reels die there. Spend the 1-credit 4K pass when the master needs it, after the motion is right.

    Do not stack a still upscaler on frame one of the clip and call it a video finish. Topaz Image and Clarity Crystal are stills rows. ByteDance is a video row. Mixing them is how you get one sharp frame and 149 untouched ones, or a clip billed as if you understood the card.

    Do not reverse the stills chain because a 4K badge is sitting on the upscaler. The badge is a ceiling, not a reason to enlarge a room. The room should already be gone.

    The whole sequence, named:

    1. Generate or shoot a still where the subject is obvious. Use text to image if you are prompting it.
    2. Cut out. Free BiRefNet, or a paid photo edit if the free matte fails.
    3. Inspect the alpha. Fix it here.
    4. Crop to the delivery ratio at native pixels.
    5. Upscale the isolated keeper.
    6. Composite onto a simple RGB plate if I2V needs one.
    7. Image-to-video on that plate.
    8. Video 4K only on the clip you will ship.

    Skip a step when the file already satisfies it. A studio-white generate may not need a cutout. A 1080p listing may not need 4K. A still that is the deliverable never needs I2V. Do not skip isolation in order to "save a pass," then spend I2V plus video key to recover.

    Run the cutout on the still you have, then the upscale, in Versely.

    FAQ

    Should I upscale first so the cutout has more pixels?

    No. A bad matte at 4K is a 4K helmet. Segmenters do not magically find hair because you paid for texture in the wallpaper. Cut out at working resolution, inspect, crop, then enlarge the keeper. If the source is genuinely tiny, recover the subject with a cutout first anyway. Upscaling the whole thumbnail just invents a larger mess for the segmenter to guess at.

    Do I feed a transparent PNG into image-to-video?

    Not by default. I2V wants a still it can treat as frame one. Alpha often becomes black, white, or a checkerboard scene you did not prompt. Composite the upscaled cutout onto a controlled backdrop, then attach that RGB plate. If a specific model has proven it respects transparency, you can skip the composite for that row. Until then, assume the plate is the scene.

    When should I remove a video background instead?

    When the file is already a clip you are keeping. Bria at 1 credit and VEED at 3 credits are footage tools. They are the right first step for a talking-head take or a product video shot on a messy desk. They are the wrong first step for a JPEG you have not isolated. For a still headed to I2V, cut the photo, upscale the keeper, then animate. Key the video only if the motion pass introduced a plate you now need to replace.